Assessing climate model accuracy and future climate change in Ghana ’ s Savannah regions

This study aimed to compare the performance of six regional climate models (RCMs) in simulating observed and projecting future climate in the Savannah zone of Ghana, in order to ﬁ nd suitable methods to improve the accuracy of climate models in the region. The study found that the accuracy of both individual RCMs and their ensemble mean improved with bias correction, but the performance of individual RCMs was dependent on location. The projected change in annual precipitation indicated a general decline in rainfall with variations based on the RCM and location. Projections under representative concentration pathway (RCP) 8.5 were larger than those under RCP 4.5. The changes in mean temperature recorded were 1 °C for the 2020s for both RCPs, 1 – 4 °C for the 2050s under both RCPs, and 1 – 4 °C under RCP 4.5, and from 2 to 8 °C for the 2080s. These ﬁ ndings will aid farmers and governments in the West African subregion in making informed decisions and planning cost-effective climate adaptation strategies to reduce the impact of climate change on the ecosystem. The study highlights the importance of accurate climate projections to reduce vulnerability to climate change and the need to improve climate models in projecting climate in the West African subregion.


INTRODUCTION
Over the past century, climate studies have intensified as scientists work tirelessly to understand climate changes and their impacts on future climate (IPCC 2007;Giménez & García-Galiano 2018).The African continent, in particular, is identified as the most vulnerable to climate change and its impacts due to the high dependence of livelihoods and various sectors of the economies on climate (IPCC 2007;Kalognomou et al. 2013;Yira et al. 2017).Consequently, irregularities in the climate negatively affect livelihoods (Boko et al. 2008;Kalognomou et al. 2013;Thornton et al. 2014), and there is a need for proper planning and implementation of sound policies to enhance sustainable adaptation and coping mechanisms.Understanding the impacts of future climate becomes an asset in managing these vulnerabilities.
In Africa, the vulnerability to climate change has been well documented, with many livelihoods and economic sectors heavily dependent on the climate.Recent studies have highlighted the impact of climate change on African agriculture and food security (Schlenker & Lobell 2010;Lobell 2013), as well as the increased risk of conflict and displacement due to climaterelated disasters.This vulnerability of the African continent to climate change poses a threat to achieving sustainable development goals.Many of the goals, such as eradicating poverty (Goal 1), achieving food security (Goal 2), ensuring good health and well-being (Goal 3), and promoting sustainable economic growth (Goal 8), are dependent on climate stability.The impacts of climate change, such as extreme weather events, sea-level rise, and changes in precipitation patterns, can lead to food shortages, displacement of populations, and increased poverty.
The projection of future climate often involves the use of climate models, which are mathematical equations that run on a three-dimensional grid (Intergovernmental Panel on Climate Change (IPCC) 2001; Rajib & Rahman 2012).General circulation models (GCMs) are one group of climate models that are often used in climate change studies, but they have a resolution of 200 km or more, which limits their ability to estimate changes in future climate at the regional scale.Either they are unable to capture mesoscale atmospheric circulations or the intensities they present are unrealistic (Christensen et al. 2007;Kendon et al. 2010;Rajib & Rahman 2012).Atmosphere-ocean GCMs (AOGCMs) are reported to produce relatively good predictions, especially on flat, noncoastal terrains (Kendon et al. 2010).However, these AOGCMs inadequately simulate the underlying West African monsoon (WAM) rainfall, which is the basis of the general rainfall pattern in West Africa (Hourdin et al. 2010;Xue et al. 2010;Gbobaniyi et al. 2014).
Regional climate models (RCMs), with a finer resolution (50 km), have been applied for West Africa and have relatively simulated the WAM climatology quite well (Afiesimama et al. 2006;Foamouhoue & Buscarlet 2006;Pal et al. 2007;Sylla et al. 2009).However, the use of single RCMs often results in some biases (Christensen et al. 2008;Teutschbein & Seibert 2012) due to the model used, the region of analysis, or the season under consideration (Nikulin et al. 2012).The use of multimodel average (ensemble mean) together with bias-correction methods has been found to enhance the performance of simulations (Teutschbein & Seibert 2010;Diallo et al. 2012;Paeth et al. 2011;Nikulin et al. 2012;Johnson & Sharma 2015).
In terms of climate modeling, recent studies have continued to focus on improving the accuracy of RCMs.For example, a study by Faye & Akinsola (2021) evaluated the performance of several RCMs in simulating the WAM and found that a multimodel ensemble approach provided the most accurate simulations.Another study by Safari et al. (2021) evaluated the performance of RCMs in simulating rainfall over Rwanda and found that a bias-correction approach improved the accuracy of the simulations.
In this study, the performance of six individual RCMs participating in the Coordinated Regional Climate Downscaling Experiment (CORDEX)-Africa project and their multimodel average (ensemble mean) were evaluated, and their performance before and after bias correction was applied to their outputs.The analysis included the comparison of monthly means with observations for both temperature and rainfall.Furthermore, these RCMs were used to project the future climate (temperature and rainfall) for the SADA zone of Ghana.

Study area
The area demarked as the Savannah zone covers the Upper West, Upper East, North East, Northern, Savannah, parts of Bono East, and Oti regions of Ghana (Figure 1).The area lies within latitudes 8°00 0 N-11°00 0 N and longitudes 0°01 0 E-3°00 0 W. The climate in the study area is primarily semiarid in the northern parts, semihumid in the central regions, and then humid in the southern parts (Liebe et al. 2009).Rainfall is characterized by significant variations, including disparities between successive seasons, and is often unpredictable in terms of onset, quantity, and coverage (Obuobie 2008;Ghansah et al. 2018).In the northern part, the mean monthly temperature falls between 27 and 36 °C, and in the southern part, it ranges from 24 to 34 °C, while relatively high evapotranspiration rates are observed with potential evapotranspiration rates exceeding mean annual rainfall amounts (Darko et al. 2018).
The primary vegetation of the SADA zone is mainly Savannah consisting of grasslands dotted with trees/shrubs such as the baobab, shea, mango, and neem.The southern parts are, however, characterized by dry forests that have a few tall trees with shrubs and grasses growing underneath (Darko et al. 2018).
A large percentage of the inhabitants in the SADA zone are peasant farmers living in scattered rural communities (Darko et al. 2018).The farmers are highly defenseless against upsets due to the limited augmentation of their income sources.Thus, any slight variation in the climate often results in poverty, hunger, and high competition for the limited available resources.Table 1 shows the description of the selected stations for the study.

Baseline data
Daily precipitation values for the baseline period of 1960-2005 for precipitation and 1975-2005 for temperature runs were obtained from the Ghana Meteorological Agency.These periods were selected because of the availability of observed data and also because the baseline period for the CORDEX-Africa projects ends in 2005.

Climate change data
RCM simulations utilized were daily precipitation and temperature values obtained from RCMs participating in the CORDEX-Africa projects.An ensemble of six RCM-GCM (Table 2) as well as their multimodel average (ensemble mean) were used in the analysis.
The ability of the models to accurately model future climate was tested by assessing their performance in simulating observed data.This test was necessary because models cannot be relied on to project future climate if they are unable to accurately simulate observed climate (Laprise et al. 2013).The representative concentration pathway (RCP) scenarios 4.5 and 8.5 of the Intergovernmental Panel on Climate Change were utilized in the projection of future climate.
Bias correction was employed using the climate model data for hydrologic modeling tool.Further details about the tool can be found in Rathjens et al. (2016).Bias correction was carried out because climate models tend to produce a biased representation of the observation (Christensen et al. 2008;Teutschbein & Seibert 2012).This is because simulations by RCMs are limited in their ability to accurately capture all the physical and chemical processes that occur in the atmosphere (Lupo & Kininmonth 2013).
The distribution mapping, which is a type of gamma method of bias correction, was chosen for both precipitation and temperature.This method works by correcting the distribution function of the simulations from the RCM to align with that of the observations (Sennikovs & Bethers 2009;Teutschbein & Seibert 2012).This method of bias correction has also proven to have high-performance ability in correcting most of the statistical characteristics of the RCM (Teutschbein & Seibert 2012, 2013;Ibebuchie 2021Ibebuchie , 2022)).

Statistical analysis
Statistical parameters, namely, root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²), were used to assess the performance of RCMS.The RMSE and MAE measure the error between the observed data and the simulated data.It is thus postulated that the smaller value (i.e., the closer it is to zero), the better or closer the stimulation value to the observed data.These parameters have been widely accepted and employed in measuring model performance in climate studies (Bennett et al. 2013;Amirabadizadeh et al. 2016).RCM outputs without bias correction and those that underwent bias correction are hereafter referred to as raw and bias corrected, respectively.
Each RCM was evaluated to determine how well they were able to simulate the past climate.The ensemble means of these RCMs were also tested to see their ability in reproducing past climatic features.

Model performance (rainfall)
RCMs simulated observed rainfall with varying levels of accuracy, as was also observed by Nikulin et al. (2012).Uncertainties in the reproduction of the observed rainfall by the raw RCMs varied across stations (Table 3).RC4A-HadGEM2 had the highest RMSE and MAE values of 86-124 and 11-70 mm, respectively, for the highest number of stations (six stations), indicative of the highest error margins and are thus said to be the least performing model.RCA4-NOAA also recorded relatively high RMSE and MAE values for the second highest number (four stations) of stations, and REMO-MPI also recorded relatively high RMSE and MAE values for three stations.
The R² of the raw RCMs ranged from 0.69 to 0.99 (Table 3).Most of the raw RCMs' R² values were higher than 0.50 for all the stations except Tamale, where all the raw RCMs recorded relatively low values ranging from 0.07 to 0.89 with REMO-MPI recording the lowest value of 0.07.The R² ranges typically from 0 to 1 and is a measure of how well the models can simulate the observed data.Values of R² closer to 1 are considered the best simulations, but values .50 are generally regarded as acceptable (Santhi et al. 2001;Van Liew et al. 2003).However, per the observed simulation and comparison with the other performance estimators, using the R² alone as a performance indicator may be misleading.
The ensemble means of the raw RCMs captured the rainfall patterns (peaks) better than most of the individual RCMs, although with varying levels of uncertainties across the stations.The RMSE and MAE for the ensemble means showed that even for the raw outputs, it performed relatively better than most individual RCMs.The values recorded ranged from 19 to 76 mm and 13 to 63 mm, respectively, for RMSE and MAE, for all RCMs at all locations (Table 3).Tamale recorded the highest values of 76 and 63 mm for both RMSE and MAE, respectively.The R² analysis showed values that ranged from 0.69 to 0.99.
The accuracy of individual RCMs improved with bias correction.RMSE and MAE values ranged from 2 to 23 mm and 2 to 22 mm, respectively.Most of the stations recorded values less than 10 mm for all bias-corrected RCMs, apart from Kete-Krachi, where all the RCMs recorded relatively high values for both RMSE and MAE, with RACMO-EARTH recording the highest value of 22 mm for RMSE.In contrast, RCA4-CanESM1 and RCA4-NOAA recorded 23 mm for MAE (Table 3).
HIRAM-NorESM1 0.66 0.12 18 12 15 11 6.9 4.8 (Continued.) R² values of the bias-corrected RCMs ranged from 0.94 to 1.00 for all RCMs at all stations (Table 3).Kete-Krachi recorded the lowest R² values (0.94) recorded by RCA4-CanESM2 and RACMO-EARTH (Table 3).As mentioned earlier, the closer the R² value is to 1 indicates a lower level of error by the model in simulating the observed data.Thus, the performance of the models in simulating the observed data is seen to improve with bias correction as was critically pointed out by Teutschbein & Seibert (2012).
Similar to the observation made for the individual RCMs, the performance of the ensemble means improved with bias correction.RMSE values ranged from 2 to 19 mm for all stations (Table 3).The highest value of 19 mm was recorded at Kete-Krachi with the rest of the stations having values less than 10 mm.MAE values obtained after bias correction also ranged from 2 to 14 mm with Kete-Krachi again recording the highest value of 14 mm.The remainder of the stations had values less than or equal to 6 mm.These indicate that the removal of biases from the simulated data improved the similarity between these and the observed data.

Simulation of monthly rainfall
The monthly simulations of rainfall patterns for each climate station per each RCM are presented in Figure 2. RCMs were mostly unable to simulate the rainfall patterns correctly.For stations with two peaks in their cycle (Bole, Kete-Karachi, and Wenchi), most of the individual RCMs either failed to detect the two peaks or simulated a third peak.For such stations, RCA-4 CanESM2 did a better job at mimicking the peaks better than the other RCMs, although with high underestimations of the rainfall amounts.For the other stations (Navrongo, Tamale, Wa, Yendi, and Zuarungu) with just one peak, REMO-MPI did better at simulating this pattern as compared to the other RCMs, albeit with some level of overestimation.RCA4-HadGEM2 was the only RCM that was unable to capture the monthly rainfall pattern present in the observed data for all the stations.
However, for stations such as Tamale and Kintampo, neither the raw RCMs nor their ensemble mean was able to capture the rainfall pattern in the observed data (Figure 2).For Tamale, all the RCMs overpredicted the rainfall from October to March and then underpredicted it from April to September.At Kintampo, on the other hand, all the raw RCMs excluding RCA4-NOAA, HIRAM-NORESM1, and RCA4-CanESM2 underpredicted the observed data for all the months (Figure 2).
The lowest underestimations in simulated data were recorded from November to March.These months fall within the dry season of the study area.On the other hand, overestimations generally occurred from April to October, when most of the stations receive rainfall.This observation can be attributed to the reported challenges faced by RCMs in simulating the WAM, especially during the rainy season.

Model performance for annual temperature
The accuracy of RCMs in reproducing mean annual minimum temperature (T min ) and maximum temperature (T max ) was higher as compared to the simulation of rainfall (Figure 3).All other RCMs apart from RCA4-HadGEM2 were able to model the observed temperature (both T min and T max ) with varying levels of accuracy.RAMO-EARTH was observed to underestimate both T min and T max at all locations consistently.At Tamale, none of the RCMs was able to model the observed pattern for both T min and T max .Barring Tamale, the MAE and RMSE values for the raw RCMs for both T min and T max ranged from 1 to 4 °C (Table 3).Values recorded for both MAE and RMSE at Tamale were relatively high for all the RCMs and ranged from 15 to 18 °C and 16 to 20 °C, respectively.R² values recorded ranged from 0.2 to 0.98, except again for Tamale where values in the range of 0.04-0.4were recorded for T min .Wenchi (T max ) and Kete-Krachi (T min and T max ) recorded some of the highest R² values (Table 3).
The ensemble means of the raw RCMs recorded MAE and RMSE values ranging from 1 to 2 °C for all stations for both T min and T max apart from Tamale that recorded MAE values of 15 and 16 °C for T min and T max , respectively, and RMSE values of 15 and 18 °C for T min and T max , respectively (Table 3).
R² values recorded for the ensemble mean were in the range of 0.04-0.98(Table 3).The lowest value was recorded for T min at Tamale, while the highest was recorded at Wenchi for T max .Similar to the observations made with precipitation, bias correction improved the ability of the RCMs in simulating the observed temperature.All the stations recorded MAE and RMSE values of 0 for all stations for both T min and T max .Since both the RMSE and MAE measure the degree of error between the observed and simulated data, it is accepted that the closer a value is to 0, the more similar the simulated value is to the observed.It indicates little or no residuals between the observed and simulated data.Thus, the MAE and RMSE values of 0 that were recorded indicate that after bias correction, the RCMs were able to simulate the observed data with a high level of performance.Also, each RCM recorded an R² value of 1 after bias correction for all stations.A further indication of the high performance of the RCMs in the stimulation of 1 implies that the simulated data are a perfect fit for the observed.The ensemble mean of the bias-corrected RCMs produced similar results to those of the individual RCMs.Zero was the value for both MAE and RMSE for all stations, as well as an R² value of 1.

Simulation of mean monthly T min and T max
Generally, T min was observed to be lowest in January and then gradually increased until it peaked in March/April.It then again declined from May until a second (smaller) peak was observed in October, after which it fell until December.For T max on the other hand, it was found to be high in January and continued to increase until it peaked in March/April and then declined sharply to its lowest in July/August, after which it increased again until December.These monthly T min and T max patterns were generally captured by the raw RCMs, albeit with varying levels of accuracy for each climate station (Figure 3).For T min , there were mixed simulations by the RCMs for the various locations.RACMO-EARTH, however, consistently underpredicted, while HIRAM-NorESM1 overpredicted for all stations.These observations are likely to be the result of the underlying boundary conditions used in the creation of these models.On the other hand, for T max , almost all the months were slightly overpredicted for most stations by almost all raw RCMs, apart from Navrongo and Zuarungu where some underpredictions were observed.RACMO-EARTH again consistently underpredicted for T max for practically all stations and months.Tamale was the only station that recorded extreme values for both T min and T max by all RCMs.
For all the other stations (except Tamale), simulations overpredicted the observed temperature by as much as 7 °C, which was recorded at Wenchi for T max by RCA4-HadGEM2 for July and August.The highest underprediction was 9 °C, registered at Navrongo for T max for April by RACMO-EARTH.For Tamale, biases of both T min and T max were extremely high.All the RCMs underpredicted both T min and T max for all the months (Figure 3).The highest underprediction was recorded in March by a value of 24 °C for T min by RACMO-EARTH, while that of T max was 28 °C recorded in February by the same RCM (Figure 3).
After correction, the biases were almost completely removed, as mentioned in the previous section.All the bias-corrected RCMs were able to reproduce the same monthly means as observed at all the stations.

Projected changes in future climate
Based on the previous results, only the bias-corrected ensemble means were considered for analysis in this section.Future projections were analyzed for each station for the three epochs according to the World Meteorological Organisation (Ondras 2011;Gulacha & Mulungu 2016), with 30-year intervals, namely the 2020s (2006-2035), 2050s (2036-2065), and 2080s (2066-2095).These epochs were then used to undertake the change analysis in future climates.The future analysis was undertaken for both RCP 4.5 and RCP 8.5 using the baseline 1960-2005 and 1975-2005 for rainfall and temperature, respectively, and each RCM and then the ensemble mean.Change in rainfall was calculated as a percentage difference, while that of temperature was derived as the absolute difference.A positive signal was an indication of an increase in the variable (either rainfall or temperature), while a negative sign indicated a reduction with respect to the baseline period.

Projected monthly rainfall
Percentage differences in rainfall between the projected rainfall amounts and the baseline amount showed variations based on location and the period under consideration (Figure 4).The 2020s, for most sites, showed either declines or the least increase in rainfall amounts as compared to the other epochs.

Projected annual rainfall
The general projections under RCP 4.5 showed a decline in rainfall amounts for all three future epochs (Figure 5).The majority of the stations recorded the highest reductions in the 2020s except for Kete-Krachi where the least decline was recorded in the 2020s and for Bole where an increase was instead observed in the 2080s.
Projections under RCP 8.5, on the hand, showed variations with respect to locations.Stations such as Kete-Krachi and Zuarungu, as well as the entire study area, had simulations similar to those projected under RCP 4.5.All the other sites had projected declines in the 2020s and increases in either the 2050s or 2080s or both.
An exception to the aforementioned observation was noted at Tamale, where projections by both RCPs for three epochs showed declines that exceeded 60%.

Projected changes in monthly temperature
Projected changes in monthly T min (Figure 6) showed that T min increased for all three epochs with the 2020s recording the least increase, while the 2080s marked the highest for both RCP 4.5 and RCP 8.5.Generally, the highest projected increases in T min were recorded in March and peaked in May, after which they declined for both RCP 4.5 and RCP 8.5.However, there were a few spikes observed in December and January for the 2050s and 2080s for both RCPs at some locations.As expected, the simulations by RCP 8.5 were generally higher than those under RCP 4.5 because of the conditions under which those scenarios were created.and November.Finally, at Zuarungu, all months recorded appreciations apart from November and December where no difference was observed.
Similar to the observations made under RCP 4.5, projections under RCP 8.5 showed that for locations such as Bole and Kete-Krachi, declines were noticed in November.Navrongo and Tamale recorded declines in November and December, while all months showed appreciations at Wa.At Wenchi, a decrease was seen in January, while no change occurred in October, November, and December.No change in T min was noticed in September and November, while a decline was recorded in October at Yendi.At Zuarungu, all months showed increases in T min apart from November when no change was recorded.
The simulated T min changes were about 1 or 2 °C for most months with a few of them recording no change (0 °C) for the 2020s.T min then increased from 2 to 6 °C in the 2050s and from 2 to 9 °C in the 2080s, with a few variations with the location.
Analogous to the projected changes in T min , the simulated T max changes were about 1 or 2 °C for most months with a few of them recording no change (0 °C) for the 2020s (Figure 6).It increased from 2 to 6 °C in the 2050s and from 2 to 9 °C in the 2080s, with a few variations with location.and in October and November for both the 2020s and 2050s.At Wa, all the months showed appreciations apart from January, February, April, and November where no change was noticed for the 2020s and March, which recorded a decline.There was also no change projected for the 2050s for April.At Yendi, all months for all three epochs showed increasing T max except March and April where T max was projected to decline, and July for the 2020s where no change was recorded.
Generally, T max is expected to increase in December, January, February, and March under the RCP 4.5 and RCP 8.5 across most of the stations (Figure 7).Comparably, T max is expected to intensify under the worst-case scenario (RCP 8.5).For instance, T max is expected to change by 6 °C at Yendi and Zuarungu under the RCP 8.5 (Figure 7).

Projected annual temperature
Projected changes in annual T min as simulated for both RCP 4.5 and RCP 8.5 are presented in Figure 8. Apart from Bole where no change was recorded for the 2020s under RCP 8.5, all other locations recorded increased T min for all future In this study, RCMs generally performed better at simulating temperature than precipitation, even without bias correction.However, bias correction is mandatory for achieving higher accuracy as RCM data tends to deviate from observed data due to their inability to integrate all the chemical and physical processes in the atmosphere (Lupo & Kininmonth 2013;Lazoglou et al. 2019).This was evidenced by the fact that after bias correction, almost all RCMs satisfactorily simulated observed data and captured rainfall peaks unique to each climate station.A similar observation had also been reported by previous studies (Sylla & Nikiema 2016;Aziz & Obuobie 2017).
The ensemble mean was observed to outperform most of the individual RCMs in most cases both before and after bias correction.Although some studies have argued that in some cases, the ensemble mean is unable to outperform the individual RCMs or GCMs, a suitable RCM or GCM that is identified could be used instead.This observation holds for some locations (Tamale and Wa) where one or two individual RCMs were observed to outperform the ensemble mean.However, the margin of error in those instances for the ensemble mean was still within the acceptable range, and therefore, the ensemble mean could be employed for further analysis.
Overall, almost all climate stations showed projected increases in rainfall amounts, especially from January to March for the 2050s and 2080s, with declines observed from August to September, which is typically the peak of the rainy season.These changes may have significant implications for agriculture, as the declines coincide with the growing season and thus may potentially lead to crop failure.
Comparing projections under RCP 4.5 and RCP 8.5 showed a similar pattern, with variations in amounts.However, RCP 8.5 projections were generally higher due to the scenario's high emissions.The response of precipitation to increasing temperatures resulting from global warming has been reported to be far from homogenous, with oscillating values ranging from À30 to 30% (IPCC 2013;Sylla & Nikiema 2016).For this study, projected changes ranged from À20 to 20% for almost all locations, with the highest decline occurring in the 2020s under both RCP 4.5 and RCP 8.5, except for Tamale.This implies that projected changes are spatially variable, and therefore, there is the need for localized adaptation strategies to suit the need of the particular location.
A trend in extreme temperatures, particularly diurnal minimum, has been observed in West Africa during the second half of the 20th century (Mouhamed et al. 2013;Darko et al. 2018).Changes in T min were generally higher than T max , with T min being the temperature component most affected by slight changes in climate (Kim et al. 2013).The resultant effect could be a persistent warmer surface temperature since the cooling of the environment depends on the lower daily temperature.Similar projections for future temperature for the West African sub-region have been made by Diallo et al. (2012) and Sylla & Nikiema (2016).Aziz & Obuobie (2017) also made comparable findings in the black Volta basin.The results from these studies suggest that the entire region is undergoing warming.
A comparison of projected changes for the entire study area and individual locations produced varied results.Although they all projected temperature increases, the degree of change varied spatially.These variations ranged from about 1 to 3 °C, but for climate change studies, a 1 °C change in temperature is considered significant and could have dire consequences for the climate (IPCC 2013).It is thus critical to consider these variations in planning and executing adaptation strategies and policies to enhance sustainable agriculture and livelihoods.
However, more RCMs are needed to explore more possibilities regarding the uncertainty of RCM parameterization.Also, additional variables need to be examined (such as potential evapotranspiration and surface runoff) for gaining insight into the influence of climate change on the hydroclimate aspects of the Savannah regions.

CONCLUSION
The accuracy of both individual RCMs and their ensemble mean in simulating observed climate improved with bias correction.However, the performance of the raw ensemble mean was better than most of the raw individual RCMs in most cases.The performance of individual RCMs varied with location.Noticeably, it was difficult for RCMs to simulate the climate pattern for Tamale for both rainfall and temperature.Although after bias correction, this situation was resolved, this challenge was again observed in the projection of future climate.These findings buttress previous studies that advocate for bias correction of RCMs before being used for further climate analysis.This is consistent with findings by Diallo et al. (2012), Gbobaniyi et al.(2014), andNikulin et al. (2012).Again, the results also support those studies that have suggested the use of the multimodel ensemble mean of several RCMs instead of using just one, since the ensemble mean was found to outperform most of the individual RCMs for most locations (Teutschbein & Seibert 2010;Nikulin et al. 2012;Gbobaniyi et al. 2014).Although there were a few instances where a single RCM performed better than the ensemble mean, generally the ensemble mean was the best-performing model.
The changes in annual precipitation for the entire study period as projected by the individual RCMs and their ensemble mean was a general decline in rainfall amounts, albeit with variations based on the RCM and location.As has been established by earlier studies, the changes in rainfall were found to be variable, oscillating between positive signals and negative signals with no particular trend.The change signal projected by the ensemble mean ranged from À20 to 20% for almost all locations, with the highest decline occurring in the 2020s.Some of the individual RCMs, however, projected change signals larger than these values.The 2050s recorded a relatively lower decline, while the 2080s generally recorded the smallest declines or in some cases increased.This observation was true for the analysis of the entire study area as well.In congruence with previous studies, simulations under RCP 8.5 were observed to be larger than the simulations under RCP 4.5.
The accuracy of simulating observed temperature by the individual RCMs improved with bias correction just as it did with rainfall, with the ensemble mean generally outperforming the individual RCMs in both cases.
The projected change in temperature showed an apparent increase in temperature over the study area for the various climate locations as well as for the entire study area.The mean temperature changes recorded were 1 °C for the 2020s and 2 °C for both the 2050s and the 2080s, in consonance with previous studies in the region as well as global predictions.

Figure 1 |
Figure 1 | Location of climate stations in the Savannah zone of Ghana.

Figure 3 |
Figure 3 | Raw and bias-corrected simulations of observed temperature (T max and T min ) from 1975 to 2005.

Figure 4 |
Figure 4 | Projected changes in monthly rainfall for the selected stations under RCP 4.5 and 8.5.

Figure 6 |
Figure 6 | Projected changes in monthly T min for the selected stations under RCP 4.5 and 8.5.

Figure 9 |
Figure 9 | Projected changes in annual T max under RCP 4.5 and 8.5.

Figure 10 |
Figure 10 | Projected changes in mean temperature.

Table 1 |
Description of climate stations used in the study Note: Latitude and longitudes in decimal degrees, elevation in meters.

Table 2 |
Details of RCMs used in this study

Table 3 |
R 2 , RMSE (mm), MAE (mm), and standard deviation of the raw and bias-corrected (in parenthesis) simulations of rainfall, maximum and minimum temperature of the individual RCMs, and their ensemble mean for each location