Distinct and combined impacts of future climate and land use change on the flow of river Rwizi in Uganda, East Africa

Globally, the threats and pressures associated with the emerging impacts of climate and land use change are acute in many low- and middle-income countries in Africa and Asia (Byers et al. 2018; Clarke et al. 2022). These nations face a complex set of circumstances that include economic constraints, competing development priorities, and inability to meet to ambitious international greenhouse gas reduction targets such as those specified in the 2015 Paris Agreement (Vairavamoorthy et al. 2008). In financial-resource-constrained and data-scarce countries particularly in the Global South, understanding quantification and assessment of the isolated impacts as well as the combined impacts of land use change and climate change on water resources is critical for resilient and sustainable management of water resources based on the best available evidence (Shrestha et al 2020; Clarke et al. 2022; Onyutha et al. 2022; Aruho Tusingwiire et al. 2023; Sempewo et al. 2023).

The current trend of increasing water demand caused due to rapid population growth, anthropogenic changes and urbanisation is expected to exacerbate the future water crisis (World Economic Forum 2017; Makarigakis & Jimenez-Cisneros 2019). The water crisis is expected to be further aggravated by emerging climate and land use change threats (Vairavamoorthy et al. 2008; Sharannya et al. 2021).

Several researchers have examined the isolated impacts of climate change on river flows for various regions in the world (e.g. Kumar 2014; Gebre et al. 2015; Zhang et al. 2019). The consensus from most studies is that the temperature will continue to increase during the next century over the world, while precipitation projections show inconsistency among regions. Previous studies that have investigated isolated impacts of both climate and land use change on river flows suggest that the resulting hydrological responses and water availability are affected either positively or negatively by these changes in various regions (Chang et al. 2015; Woldesenbet et al. 2017; Yang et al. 2017; Gebrechorkos et al. 2023). More recent studies (Dong et al. 2014; Zhang et al. 2016; Yin et al. 2017) have investigated the combined impacts of land use and climate change on river flows. Some studies have reported that river flows were more sensitive to land use change than to climate change (e.g. Dong et al. 2014; Yin et al. 2017), while others reported otherwise (e.g. Guo et al. 2008; Shrestha & Htut 2016; Ross et al. 2021). These contradicting results could be attributed to the intensity and types of the land use changes, river catchment characteristics or climate variations of the catchments. The discrepancies to a degree could also be attributed to uncertainties arising from available data and the structure and parameters of the models (Global Climate Models (GCMs) and hydrology model) used in these studies (López-Moreno et al. 2011).

It is, therefore, evident that there is still limited knowledge on the relative influence of climate and land use change on river flows (Howells et al. 2013). This is of great concern for water resource management with respect to mitigating and adapting to future changes in land use and climate changes using the best available evidence. Therefore, catchment-specific studies are required to provide comprehensive and accurate information for water resource management. This study investigated the water-stressed river Rwizi catchment in western Uganda, which is currently the sole source of water supply to Mbarara city, the second largest in Uganda. The river has experienced a reduction in flows over the years which existing studies have attributed to climate change (e.g. Atim 2010; Nyeko-Ogiramoi et al. 2010; Wanyama 2012). Other studies in the catchment attributed the changes to land use and land cover (LULC) changes (e.g. Nseka et al. 2022). A recent study by Onyutha et al. (2021) attempted to provide the distinctive impacts of land use and climate change on the river flows with the study attributing 73% of the flow variations to climate change and 17.1% being attributed to other factors. However, this study did not consider the effect of future climate and land use changes on river flows.

There is, therefore, a lack of knowledge on the driving factors that could cause future changes in the river Rwizi's flow and the region. In this study, the distinct and combined impacts of climate and land use change on river Rwizi flows was investigated using an ensemble of six CORDEX Regional Climate Model (RCM) results and Cellular Automata–based Artificial Neural Network (CA-ANN). Simulations of current and future river flows were undertaken using the Soil and Water Assessment Tool (SWAT) model.

The main objective of this paper is to assess the distinct and combined impacts of future climate and land use change on river Rwizi flows using a combination of statistical downscaling techniques, machine learning and physically based hydrological modelling techniques. Details of the methodology are presented in the following sections.

The description of the datasets used this study is presented in Tables 1 and 2.

To investigate the impacts of climate change and LULC on hydrological processes in the upper Rwizi catchment, land use, topographic and soil and hydro-meteorological data were obtained. A digital elevation model (DEM) of 30 m resolution for the year 2009 was acquired from the United States Geological Survey – Shuttle Radar Topography Mission (USGS-SRTM) in geo-tiff format. The soil map (1: 5,000,000) developed by the Food and Agriculture Organization of the United Nations (FAO-UNESCO) was downloaded from http://www.fao.org/soils-portal/soil-survey/soil-maps-and-databases/faounesco-soil-map-of-the-world/en/.

Landsat satellite images for the years 1984, 2000 and 2016 were acquired from the United States Geological Survey (USGS) website (https://earthexplorer.usgs.gov/) at a 30 m spatial resolution and classified into LULC types of farmland, bushland, open water, woodland, built-up area, wetland and forested area using the maximum likelihood supervised classification tool in ArcGIS following the generalised classification system (Atay Kaya & Kut Görgün 2020; Palmate et al. 2022). The land use data were in raster format. Land use properties were sourced directly from the SWAT model database. The observed discharge data for the only existing gauged station for River Rwizi Mbarara Water Works (No. D81224) hydrological station for the river Rwizi was acquired from the Directorate of Water Resources Management (DWRM), Ministry of Water and Environment, Uganda, and used for model calibration and validation.

Due to large data gaps (up to 10–15 years) in the observed rainfall data acquired, reanalysis data were used (Morales-Moraga et al. 2019; Nakkazi et al. 2022). This dataset is readily available to the public domain and therefore widely applied in data-scarce regions. Reanalysis data were bias-corrected and evaluated for suitability using the method proposed by Berhanu et al. (2016). A comparison was made between observed rainfall data obtained from the Uganda National Meteorological Authority for the period (2000–2009) with five reanalysis datasets, i.e. AgMERRA, CRSR, CHIRPS, ERA5 and PRINCETON, to determine which data represented the observed catchment rainfall best. The AgMERRA reanalysis dataset that provides daily, relatively high-resolution (0.250° × 0.250°), continuous, meteorological series over the period 1980–2010 achieved a Pearson correlation of 0.93 thereby representing the catchment rainfall best. Therefore, this dataset was adopted for further analysis in this study.

There are a vast number of climate models available in the public domain for providing climate change projections (Dosdogru et al. 2020; Atay Kaya & Kut Görgün 2020). However, for this study, outputs from CORDEX RCMs were adopted. CORDEX RCMs were chosen for this study because of their higher spatial resolutions (0.44° × 0.44°) as compared with those of the CMIP3 and CMIP5 GCMs. In addition, according to the studies by Onyutha et al. (2019) and Kisembe et al. (2019) on how well the climate models reproduce observed rainfall variability, CORDEX RCMs reproduced variability in daily rainfall over the study area better than the CMIP3 and CMIP5 GCMs. A total of six CORDEX RCMs were adopted based on previous research by Onyutha et al. (2019) as well as the availability of variables from the data portal.

Simulated climate change output variables of both temperature and precipitation for the CORDEX RCMs for both mid-range RCP4.5 and extreme RCP8.5 scenarios were obtained from the website https://esgf-index1.ceda.ac.uk/search/esgf-ceda/. To prepare the simulated climate data for catchment-based hydrologic impact analysis, the data were bias-corrected using the distribution mapping technique (Teutschbein & Seibert 2012; Alramlawi & Fıstıkoğlu 2022). The CMhyd (climate model data for hydrologic modelling) tool developed by Rathjens et al. (2016) was used.

The procedures of projecting future land use change involved three steps: (1) establishing the transition matrix, (2) generating suitability maps and (3) projection of the LULC map. The procedures for projecting land use and land cover change (LULCC) were undertaken using the open-source Module of Land Use Change Evaluation (MOLUSCE) tool for QGIS 2.8.10 version with the CA model, following the procedures of Jogun et al. (2019). A combination of CA-ANN was found to be one of the most efficient methods for simulating complex and non-linear LULC which are also specific to the research area (Jogun et al. 2019). The plugin takes input data from categories of land use and explanatory variables, trains the LULC model using input data and well-known algorithms, forecasts potential LULC based on training input and validates findings based on past reference data (GIS LAB 2014).

For modelling LULC transition probability, ANN multi-layer perception was adopted because it offers the highest accuracy compared with the other algorithms such as logistic regression, multi-criteria analysis (MCA) and weight-of-evidence methods (Jogun et al. 2019). The method was also adopted because it is suitable for data-scarce situations and for capturing complex non-linear features in modelling processes (Li & Yeh 2002).

Then, the generation of suitability maps of the potential driving factors of the LULC transitions was developed for each LULC class using MCA. A simulated land use map of 2048 was produced based on a Monte Carlo CA model approach (Pijanowski et al. 2002). Previous research has outlined the key elements that influence potential land use changes (Park et al. 2011; Jokar Arsanjani et al. 2013; Musa et al. 2016; Shafizadeh-Moghadam et al. 2017). Slope, distance to roadways and urban centres, and population density were the primary variables used in land use change simulation modelling to generate suitability maps. Finally, the projection of the LULC maps was done based on the transition probability matrix and the generation of suitability maps.

Pearson's correlation function of the MOLUSCE plugin in QGIS was used to check the correlation among the spatial–temporal changes and compute the LULC changes and category of each area between 1984 and 2016 to generate two LULC change maps (1984–2000 and 2000–2016). The purpose of this stage was to compute the changes in area in km² between the LULCs of 1984 and 2016.

After achieving satisfactory validation results, the simulated LULC multi-resolution analysis for the 2040s (i.e. 2048) was undertaken to determine the difference between the reference map and the simulated map in position and quantity (Pontius & Suedmeyer 2004). The analysis was undertaken at different spatial resolutions (Jogun et al. 2019). Details of the methods are described in the studies by Costanza (1989) and Castella & Verburg (2007). The approach was adopted because it has been widely applied in the literature (Lin et al. 2011).

Rainfall–runoff modelling was performed using the SWAT extension in ArcGIS 10.2.1 version. The SWAT model is a physically based, semi-distributed model that can predict the movement of water in complex catchments with varying soils, land use and management conditions over long-term periods (Arnold et al. 1998; Neitsch et al. 2009). The SWAT model was used because it is one of the most widely used hydrological models in the world (Gassman et al. 2010) and is widely applied in data-scarce regions, particularly in middle- and low-income countries (Nakkazi et al. 2022; Sempewo et al. 2023). Furthermore, the model was applied because of its modularity capabilities, computational efficiency, the capacity to predict long-term effects as a continuous model and the ability to use easily accessible global datasets

Usually, not all parameters of the model have the same effect on stream flow (Abbaspour et al. 2017). Therefore, sensitivity analysis was done with the semi-automated Sequential Uncertainty Fitting (SUFI-2) algorithm using the global sensitivity method. Based on the significance of the ranked values, the t-statistics and p-values of the parameters were used to rank the various parameters considered to affect flow (Arnold et al. 2012).

Calibration and validation were done through the semi-automated SUFI-2 algorithm within SWAT Calibration and Uncertainty Procedures (SWAT-CUP) version 5.2.1. For catchments in the sub-tropical and tropical zones like the study area, the SUFI-2 method is the technique that shows the most promise for calibration and uncertainty (Khoi & Thom 2015). In addition, SUFI-2 has a higher computational efficiency than other methods (e.g. ParaSol, Particle Swarm Optimization (PSO) and Generalized Likelihood Uncertainty Estimation (GLUE)) requiring a sample size for one iteration ranging from 500 to 1,000 runs (Khoi & Thom 2015).

The SWAT model was calibrated by comparing the simulated flow and the historical flows of the only gauging station within the catchment from the Mbarara Water Works gauging station D81224 (see Figure 1) for a period from 1 January 1982 to 31 December 1989. The validation of the model was based on the period starting from 1 January 1990 to 31 December 1995 to test if the calibrated parameters adequately represent the catchment processes.

The calibrated SWAT rainfall–runoff model was used to evaluate the temporal variation in runoff to assess the impacts of future land use and climate change on the upper Rwizi catchment behaviour. Different scenarios were used for analysing the impact of climate and land use changes in the future period. For each scenario, the changes in flow were calculated by comparing stream flow between the reference period and simulated flow (average of all the RCMs) future period.

The change comparison was based on the ten-year extreme quantiles. The high- and low-flow events were extracted as peak over thresholds (POTs) using the frequency analysis considering the non-stationarity tool (Onyutha 2017a, 2017b). The independence criteria for the extraction of the POT were followed using studies by Lang et al. (1999) and Onyutha (2017b). The POTs of observed and future flows were compared on a return period basis in the context of frequency analysis to estimate the impact of climate change, land use and their combined effects on flow.

• (a)

  Annual flows

Details of the approach can be found in the study by Tomer & Schilling (2009). However, the approach is only suitable for regions where precipitation matches evaporative demand and has been criticised for not being applicable to all hydroclimatic conditions (Renner et al. 2014). To address this limitation, the expanded framework of Marhaento et al. (2017) was adopted.

• (b)

  High and low flows

In order to distinguish the impacts of future climate change and land use change on hydrological extremes, the study followed the methodology of Chen et al. (2019). In this regard, four scenarios (Table S3, Supplementary Material) were considered, with a reference scenario and altered scenarios as per the tables. Using the SWAT model, the runoff series for each scenario were obtained and the corresponding extreme hydrological indicators were derived.

Table 3 presents the results on the future monthly changes in rainfall. The results show that the average annual precipitation in the future is expected to decrease by 11.68% (10.02 mm) for RCP4.5 and 12.2% (10.47 mm) for RCP8.5. This indicates that the RCP8.5 will be drier than the RCP4.5. Nevertheless, the mean precipitation for the future period in the June, July, August and September (JJAS) season will increase by around 5% (13.24 mm) and 1% (3.27 mm) for the RCP4.5 and RCP8.5 scenarios, respectively. On the other hand, the mean precipitation for the future period in the January–February (JF) season will decrease by around 45% (51.62 mm) and 48% (54.48 mm) for the RCP4.5 and RCP8.5 scenarios, respectively. This indicates that on average under both RCP4.5 and RCP8.5 scenarios, long dry spells are to be expected in the future.

During the March, April and May (MAM) wet season, a decrease in rainfall of about 18.15% (55.21 mm) and 16.76% (50.98 mm) for RCP4.5 and RCP8.5, respectively, was projected for the future period. Similarly, for the October, November and December (OND) wet season, a decrease in rainfall of about 7.5% (26.75 mm) and 6.57% (23.43 mm) for the RCP4.5 and RCP8.5 scenarios, respectively, was projected for the future period. This indicates that there will be less rainfall in the wet season in the future, as compared with the baseline period for both the RCP4.5 and RCP8.5 scenarios.

However, the months from August to October experienced an increase in rainfall compared with the baseline period for both RCP4.5 and RCP8.5 scenarios.

The results show that there was an increase in T_max within the Rwizi catchment under both the RCP4.5 and RCP8.5 scenarios (Table 4). The projected increase in T_max was about 0.9 and 1.0 °C for the RCP4.5 and RCP8.5 scenarios, respectively. The season of JJAS and JF had the highest increase in T_max, while the lowest increase in T_max was projected in the MAM and OND seasons for both scenarios.

According to the findings, there was a general increase in mean annual T_min within the Rwizi catchment under both the RCP4.5 and RCP8.5 scenarios. The projected increase in mean annual T_min was about 0.7 and 0.8 °C for RCP4.5 and RCP8.5 scenarios, respectively. On a monthly basis, all months were observed to have had an increase in the average monthly T_min for both scenarios, i.e. RCP4.5 and RCP8.5. In addition, the season of JJAS had the highest increase in T_min of about 0.8 and 0.9 °C under the RCP4.5 and RCP8.5 scenarios, respectively.

This means that the dry season will be hotter than in the baseline period. This can be explained from the fact that JF and JJAS are dry seasons while MAM and OND are wet seasons, and the dry season usually experiences higher temperatures than the wet seasons. The results also show that JF season will be drier than the JJAS season under both RCP4.5 and RCP8.5 scenarios.

The results for land use change analysis are presented in Table 5.

The results show that compared with the other classes, the farmland land use class had the highest increase in area with 90.7 km² (74.3%) from 1984 to 2000, 302.1 km² (142%) from 2000 to 2016, and 1.54 km² (0%) from 2016 to 2048 (Figure 3(a), Table 5). There was also a high rise in the built-up area class with a total of 27.5 km² (55%) from 1984 to 2000, 252.3 km² (326.8%) from 2000 to 2016, and 121 km² (36.7%) from 2016 to 2048. On the other hand, bushland had the highest decreases in LULC of −29.9 km² (−2.4%) from 1984 to 2000, −301.7 km² (−25.2%) from 2000 to 2016, and −110 km² (−12%) from 2016 to 2048.

Open water and wetland marginally decreased over the period 1984–2016 by −8.5 km² (−0.41%) and −15 km² (−0.7%), respectively. For the period 2016–2048, wetlands are also expected to marginally decrease by 2 km² (−0.6%). Many people in the region have encroached on the banks of the river Rwizi and other wetlands, which could explain the reduction in the area covered by open water and wetlands.

Another significant decrease was in the forest cover with a loss of −191.6 km² (−90.2%) in the period 1984–2016. From 2016 to 2048, the predicted loss is about −8 km² (−39%). Woodland class also decreased by 126.4 km² (97.5%) during the period 1984–2016. The decrease in both woodland and forested area could be explained by the fact that on average the population in the Rwizi catchment increased by 2.9% annually in the period 1984 to 2016. With the increase in built-up area and farmland, more land had been cleared for agriculture and settlement (Kuloba 2017). In addition, in the upper Rwizi catchment, over 98% of households use firewood and charcoal for cooking (NEMA 2004), which could also explain the decrease in both woodland and forested areas within the catchment over the study period.

Table 6 presents the final kappa statistics, while the percentage correctness of simulated land use of 2016 in reproducing reference/classified 2016 land cover during model validation is presented in Table S4 in the Supplementary Material. An overall kappa value of 0.524 and 66.86% correctness estimated by MOLUSCE indicate good ability of the model in projecting future LUCL of the catchment according to Landis & Koch (1977).

The results of the sensitivity analysis for the Rwizi River are presented in Table S1 in the Supplementary Material. The results show that out of the 17 flow parameters, only nine were the most sensitive parameters, and these were further used in the calibration process of the model output. The most sensitive parameter was channel effective hydraulic conductivity (CH_K2) in the Rwizi River, which is in agreement with the previous study by Anaba et al. (2017). The rankings of the flow parameters are presented in Table S1 in the Supplementary Material.

Table 8 shows the results of annual runoff for different scenarios under RCP4.5 and RCP8.5 during the period 2021–2050.

The results show that the combined impact of land use change and climate change resulted in a 26.92% and 20.95% increase in annual runoff for the RCP4.5 and RCP8.5 scenarios, respectively, for the average values of all the six RCM models, during the period 2021–2050. The land use change scenario alone showed an increase in the total annual runoff by 47.3% during this period. The influence of land use changes on total annual runoff was more dominant than that of climate change. This could be attributed mainly to the increase in built-up area between the baseline period (2000) and the future period (2048), which would lead to an increase in impervious surfaces, reducing infiltration and increasing surface runoff. During this period, there was drastic land use change that occurred in the catchment over the baseline and future periods. Forest land declined drastically from 9.2% coverage in the baseline period to 0.6% coverage, while built-up area rose from 3.7% to 21.6% of the catchment area, which in turn increases the stream flow during the rainy periods due to increased impermeable surfaces in urban areas/pavements (Anaba et al. 2017). These results are also in agreement with assertions by both Tang et al. (2005) and Nie et al. (2011) that increase in built-up areas increases impermeable surfaces, and diminishing forest land speeds up surface runoff. These results are also in agreement with a study by Baker & Miller (2013), who assessed the impacts of land use change on water resources using the SWAT model for the Njoro watershed in Kenya's Rift Valley and reported that land use changes resulted in increased surface runoff. However, the climate change alone scenario led to a reduction in annual runoff by 4.34% and 9.8% under RCP4.5 and RCP8.5, respectively. This decrease in annual runoff can be attributed to a decrease in annual rainfall, which decreased by 11.68% (10.02 mm) and 12.2% (10.47 mm) under RCP4.5 and RCP8.5, respectively.

The results also show that LULC and climate change both affect future high flows based on ten-year quantiles indicating a risk of impacting frequent hydrological flood events in the catchment. The average projected increase in high-flow events was 64.7% (22.9 m³/s) and 73.5% (25.9 m³/s) in RCP4.5 and RCP8.5, respectively, for the climate change scenarios in contrast to a projected increase of about 6% (2.03 m³/s) in high-flow quantiles for the land use change scenario. On the other hand, the combined (land use and climate) change scenarios resulted in an average estimated rise of about 77.7% (27.45 m³/s) and 75.05% (26.5 m³/s) for the ten-year high event in RCP4.5 and RCP8.5, respectively. For the high-flow quartiles, LULC had the least impact in contrast to climate change, which had the highest impact.

The average projected increase in low-flow events for the ten-year low event was approximately 46.18% (0.63 m³/s) and 45.92% (0.626 m³/s) for RCP4.5 and RCP8.5, respectively, under the climate change scenarios. For the land use change scenario, a projected decrease of about 32.69% (0.45 m³/s) in low-flow quantiles for the ten-year low event was observed. The combined (land use and climate) change scenarios resulted in an average estimated rise of 44.37% (0.605 m³/s) and 45.79% (0.624 m³/s) for the ten-year low event in RCP4.5 and RCP8.5, respectively.

Contrary to the findings of this study, Onyutha et al. (2021) attributed the changes in historical flow to climate variability in the upper Rwizi catchment. However, Zhang et al. (2016) reported that climate change has a greater impact than land use change in both historical and future periods, although they cautioned that the effects of land use change might be significant if the climate gets drier in the future. It is indeed true, as this study had reported, that there is a decrease in precipitation in the future for both RCP4.5 and RCP8.5, which means the future could be drier as per these results. This could explain the reason for land use having a more pronounced impact on annual flows that climate changes, as reported by Zhang et al. (2016).

Furthermore, the attribution analysis was carried out across the six RCM models to cater for uncertainty across the models. This was done for the combined scenario to try and understand the contribution of future land use and climate change of future flow variations. The highest contribution by climate change under RCP4.5 of about 31% was recorded by 2 = RCA4-IPSL-IPSL-CM5A, 3 = RCA4-CNRM-CERFACS-CNRM-CM5 and 4 = RCA4-NOAA-GFDL-GFDL-ESM2M. Under the RCP8.5 scenario, 1 = CCLM4-8-17-CNRM-CERFACS-CNRM-CM5, 2 = RCA4-IPSL-IPSL-CM5A, 3 = RCA4-CNRM-CERFACS-CNRM-CM5 and 5 = REMO2009-ICHEC-EC-EARTH also contributed over 31% to the changes in future flow changes. The climate change contribution to the future flow variations for all the models under both RCP4.5 and RCP8.5 ranged between 25% and 31%. It should also be noted that the contributions of the climate changes to flow variations under RCP8.5 were on average either equal to or higher than those under RCP4.5 as shown in Table 9.

The isolated impact of land use and climate change on magnitudes of high flow is presented in Figure 8(b). The results show that climate change contributed to 8% and 7%, while land use change contributed to 92% and 93% of the increase in the total magnitude of flow under RCP4.5 and RCP8.5, respectively. The results indicate that climate change is the dominant factor contributing to changes in future extreme flow quantiles within the Rwizi catchment. The increase in the magnitude of high-flow quantiles can be attributed to the increase in rainfall and the decrease in the bushland area, favouring arable land and built-up areas in the catchment.

The isolated impact of land use and climate change on magnitudes of high flow is presented in Figure 8(c). The results show that climate change contributed to 0.5% and 0.4% (almost zero), while land use change contributed to 99.5% and 99.6% (almost 100%) under RCP4.5 and RCP8.5, respectively, of the increase in the magnitude of future low-flow quantiles. It can be observed that all regional climate model runs project an increase in low-flow quantiles with increasing return periods.

The results indicated that the future changes in low flows were majorly influenced by land use changes with the changes in low flow attribution being about 99.5% and 99.6% for RCP4.5 and RCP8.5, respectively, which is consistent with the study conducted for the Gilgay Abay River basin by Dile et al. (2013), who revealed that climate change appears to have an insignificant impact on low-flow conditions of the river basin. The land cover impacts are like those of high flow. A reduction in grassland and forest area helps alleviate the duration and severity of low flows in both basins. Furthermore, land cover changes are proportionate to changes in low flows.

The study results suggest that PRINCETON, ERA5 and CFSR reanalysis data tended to overestimate the mean monthly precipitation, while the CHIRPS reanalysis data underestimated the mean monthly precipitation for the Mbarara rain gauge station. Based on the study results, it is suggested that in the absence of in situ observed rainfall data, AGMERRA reanalysis data provide the best dataset for use in studies in catchments within the same climatic zone and with precipitation patterns similar to the Rwizi catchment.

The results of this study are in line with the CORDEX RCM study findings by Onyutha (2020), which suggested that the future annual amount of rainfall in the study region would decrease. The results are also in line with the study by Onyutha et al. (2019), which that utilised the CMIP3, CMIP5 and CORDEX RCMs to project a decrease in future rainfall for the study area using both RCP4.5 and RCP8.5 scenarios.

However, some studies have reported an increase in rainfall contrary to the findings of this study within the study area, e.g. Nyeko-Ogiramoi et al. (2012) and Nyeko-Ogiramoi (2011), which projected an increase in precipitation in the Rwizi catchment over the future period. In another study, Nyenje & Batelaan (2009) projected an increase in precipitation in the nearby Sezibwa catchment using the HadCM3 GCM. The difference could be attributed to the fact that in both these studies, GCM outputs were used for climate change projection, which are on a coarser resolution than the CORDEX RCMs. This is indeed true as Onyutha et al. (2019) observed that climate change conditions varied from one set of climate models to another, which they attributed to the spatial resolution in the models. This is because the coarser resolution of the GCM models causes the GCMs to smoothly reflect the orography, thus reducing the ability to depict the effect of topographical characteristics on rainfall variability across regions (Onyutha et al. 2019).

The findings of this study are in agreement with earlier findings from Jassogne et al. (2013), who articulated an increasing trend in minimum and maximum temperature over the next 50 years beginning from 2015. Diffenbaugh & Scherer (2011) also asserted that if there were a continued rise in carbon dioxide concentrations, a permanent heat regime over the next four decades would be experienced in different locations over the globe. Consequently, the increase and decrease in T_max and T_min would mean an increase in the difference between T_max and T_min, and thus an increase in evaporation (Nyeko-Ogiramoi 2011). An increase in evapotranspiration would mean a reduction in the stream flow for the case study catchment.

The study results are in agreement with Barasa et al. (2010), who established that the highest increases in land use systems were in agriculture and grasslands, while the highest losses were in degraded grasslands and woodland/forest in Uganda. Also, Wanyama (2012) reported a decrease in grassland and forest area of 17,529 ha (−17.7%) and −1,406 ha (−39.1%), respectively, between the period 1954–2010, which is in close agreement with the findings of this study. The study results suggest the need for focused efforts on afforestation and wetland protection as suitable adaptation measures for the Rwizi catchment. Furthermore, the study results suggest the need for alternative cooking-fuel options to reduce the demand for firewood and charcoal in the Rwizi catchment.

The study results suggest that for the case study catchment, the influence of land use change was more predominant when compared with climate change. These results are in agreement with studies by Piras et al. (2014) and Islam et al. (2014), who both reported that climate change led to a reduction in mean annual rainfall which resulted in a decline in mean annual runoff. In addition, according to the findings, the combined effects of land use and climate change on changes in flow are significantly different from the effects of each factor alone. This highlights the need to consider the effects of both land use change and climate change when planning, designing and managing water resource systems.

The results agree with the findings of Dosdogru et al. (2020), who found that LULC change did not have a substantial impact on peak stream flows in the Upper Cahaba River catchment. More impact on the high flows will, however, further compound the soil erosion situation and increase sediment transport, making stream-flow muddy and impacting aquatic life and hence downgrading ecosystem services (Dosdogru et al. 2020).

The results show that LULC and climate change both affect future low flows, indicating a risk of impacting frequent hydrological droughts in the catchment. The projections suggest that these impacts are compounded with the increase in return periods.

The results suggest that the projected decrease in future precipitation for both RCP4.5 and RCP8.5 scenarios will amplify the effects of land use change on annual average river flows. However, the study results also suggest that climate change will have a greater impact on peak flows when compared with land use change. The study findings are affirmed in a previous study by Hung et al. (2020), which concluded that climate change has a greater impact on peak flows compared with land use change. Another study by Dibaba et al. (2020) in highland Ethiopia also found that climate change had a predominant influence on stream flow extremes compared with LULC effects. The percentage increases in high-flow quantiles caused by the combined land use and climate change scenarios were higher than the increases observed in individual scenarios of land use change and climate change.

Although LULC has also impacted the duration and severity of high flow, their significance is affected by the amount of land cover change. Many recent scholars have postulated that land use change increases peak flow and runoff because of decreased infiltration capabilities of the soils (Anaba et al. 2017). On the other hand, reduction in grassland and forest area produces shorter duration and more severe high flows. These results highlight the significance of considering both land use change and climate change together when analysing the impacts on high-flow events.

On the other hand, the low-flow study findings emphasise the importance of considering both climate change and land use change when assessing the impacts on low-flow events. The contribution of land use change to the increase in low flows can be attributed to the large, combined area of grassland and forest in the catchment during the projection period. In addition, it could also be attributed to the increase in population and hence settlement, which further leads to an increase in impervious surfaces through urbanisation and infrastructure development like roads, buildings and pavements. These surfaces would prevent water from infiltrating into the ground and as a result lead to reduced groundwater recharge during the rainy periods. This would result in reduced baseflow contribution to the river in the dry spells.

This aim of the study was to undertake a comprehensive assessment of the distinct and combined impacts of future climate and land use change on river Rwizi flows using a combination of statistical downscaling, machine learning and physically based hydrological modelling techniques.

The assessment was undertaken using the physically based semi-distributed SWAT model using three scenarios, i.e. LULC change, climate change and combined (climate and LULC change). The study also applied an ensemble of six CORDEX RCM climate models under the moderate (RCP4.5) and high (RCP8.5) emission scenarios to assess changes in future climate for 2050. The 2050 future land use was projected using a CA-ANN.

This study was driven by the desire to provide evidence-based insights into the key factors responsible for the changing hydrology of water-stressed catchments with limited observed data. The findings from such assessments can act as a benchmark for government ministries responsible for water resources management, river basin organisations and other non-governmental organisations to make decisions on water regulation, water conservation and regulation of plans for hydraulic infrastructure (Bai et al. 2019; Onyutha et al. 2021; Bahati et al. 2021).

The results demonstrated that land use change and climate change as well as their interactions place substantial pressures on river flow. More specifically, the results showed that changes in future annual flow and low flow were attributable mainly to land use change, while future high flows were attributed to climate change. The isolated impacts of land use change and the combined impacts showed an increase in future total annual river flows of 47.3% and 20.9%, respectively. However, the isolated impacts of climate change showed a reduction in future total annual flow of 4.34% and 9.8% under RCP4.5 and RCP8.5, respectively. The influence of land use changes on total annual runoff was more dominant than that of climate change. The findings show that LULC change and climate change affect the duration and severity for both high and low flows. Specifically, the more the land cover changes, the more the duration and severity of high- and low-flow events will be affected.

Land use change alters the composition and configuration of river catchments hence impacting the river catchment ecosystems (Bai et al. 2019; Onyutha et al. 2021; Bahati et al. 2021). Studies have shown that land use significantly impacts river flow through changes in pervious surfaces/runoff coefficients (e.g. Anaba et al. 2017). Spatial changes in land use over time have significant impacts on hydrological risks and the capacity of catchments to provide ecosystem services (Jogun et al. 2019). Scholars have recently used LULC change as a proxy for assessing impacts on river catchment ecosystems (Paudyal et al. 2019). Climate change is equally a key driver impacting hydrological flood and drought risks (Onyutha et al. 2021).

Similar to studies elsewhere, climate change has been found to be responsible for escalating the frequency and severity of extreme events, such as temperature and precipitation patterns. Climate change affects river catchment ecosystems and flows by modifying the biophysical processes, which in turn pushes the flows to the extremes. This is expected to increasingly threaten river flow and ecosystems and is expected to become a severe threat in the future (Nakkazi et al. 2022).

Although previous studies agree that LULC and climate change impact hydrological flow and extreme events, it should be noted that the impacts can produce varied hydrologic responses for the same catchment under different geographic and climatic conditions due to data variability and limitations. While bias correction has been adopted to overcome data limitations, further investigations will be required particularly when more data become available. Despite the limitations in this study, the results have methodological, practical and policy applications, which support the use of hydrological modelling to inform allocation, prioritisation and decision-making in a bid to accelerate progress towards the United Nations Sustainable Development Goal (SDG) target number 6.4–6.6. The study also supports the hypothesis that a combined scenario analysis is key to evaluating effects of driving factors on future hydrological changes in river basins. These tools facilitate a better understanding of the change impacts and projections on river basins, enabling the generation of tailored catchment-specific policy suggestions. However, it should be noted that impacts will vary from location to location (Bai et al. 2019; Qu et al. 2020). While this analysis has been applied to the Rwizi catchment, the same approach can be applied to other water-stressed catchments to come up with location-specific results and mitigation strategies like planting trees and restoring forests to enhance water retention, regulate stream flow and mitigate the impacts of climate change since trees play a crucial role in maintaining hydrological balance by absorbing and releasing water through transpiration (e.g. Azarnivand et al. 2020; Meaza et al. 2022; Kingsbury-Smith et al. 2023). In addition, implementing sustainable agricultural and land use practices, such as contour ploughing, cover cropping and reduced tillage, can minimise soil erosion and improve water retention (Kajembe et al. 2005; Nyssen et al. 2010; Sarvade et al. 2019; Telles et al. 2019; Rajbanshi et al. 2023). This helps maintain a more stable stream flow pattern (Näschen et al. 2019).

Finally, planning for water resources management in the Rwizi catchment should consider both land use change and climate impact as they are important in assessing the adequacy of water management strategies. While designing management strategies to prevent future extreme events like flood and drought, climate change deserves more attention. Pragmatic strategies for sustainable water and environmental resources in the catchment should be developed.

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

The authors declare there is no conflict.