Effects of climate and anthropogenic changes on current and future variability in flows in the So'o River Basin (south of Cameroon)

Variations in river flow generally result from interactions between climate change and/or anthropogenic changes (Diem et al. 2018; Oudin et al. 2018; Ebodé et al. 2021a). However, the study of the impact of these environmental forcings is done in different ways. In some works, this impact has been studied using the historical/contemporary period (Cissé et al. 2014; Bodian et al. 2020; Descroix et al. 2020). Meanwhile, other studies have addressed it over future periods to plan how available water resources will be distributed among different sectors (Abe et al. 2018; Fentaw et al. 2018; Dibaba et al. 2020). Until now, there are few studies (Ardoin-Bardin 2004; Sighomnou 2004) that have addressed both the study of the impact of environmental forcings on flows using current and future climate projections in some regions like Central Africa, due to the scarcity of data. The active network of this part of the continent now only includes 35 stations in the DRC (i.e. one station for 67,000 km²), 22 in Cameroon and less than 20 for all of Gabon, Congo and Central African Republic (Bigot et al. 2016). Such studies are urgently needed to assess the current and future evolution of water resources in a given region, in order to put in place different water management strategies.

In equatorial Central Africa, recent studies focusing on the impact of environmental forcings on flows during the historical period have highlighted a drop in average and extreme flows linked to a drop in rainfall for some basins such as the Ntem and the Ogooue (Conway et al. 2009; Ebodé et al. 2020; Ebodé et al. 2021b). In other basins such as the Nyong and the Mefou, it has been demonstrated that maximum flows may stay the same or increase, as a result of changes in land use patterns. These changes, which are essentially marked by an increase in impervious areas (buildings, roads and bare soil) to the detriment of the forest, would then compensate for the rainfall deficit by an increase in runoff, hence maximum flows will stay the same in the larger basins and increase in the smaller ones (Ebodé et al. 2022). In forested equatorial Central Africa, there are few studies addressing the impact of environmental forcings on flows during future periods using state-of-the-art methodology and new data (Akoko et al. 2021).

In other regions of the world, predictive hydrological modeling is done using a global or distributed/semi-distributed approach. The global approach considers the watershed as a single entity. The GR2J (Rural Engineering model with two daily parameters) and GR4J (Rural Engineering model with four daily parameters) models are some of the reference models generally used in this type of approach (Perrin et al. 2003; Bodian et al. 2012; Amoussou 2015). These models take into account parameters such as precipitation, evapotranspiration and soil water capacity. In the distributed/semi-distributed approach, on the other hand, the watershed is considered as a complex entity and flow modeling requires a subdivision into homogeneous elementary surfaces (Legesse et al. 2003; Githui et al. 2009; Taleb et al. 2019). Distributed/semi-distributed models require a wide range of input data, ranging from physical characteristics of the basin (slopes, land cover, soils, etc.) to meteorological data (precipitation, maximum and minimum temperatures, wind speed, relative humidity, insolation, etc.).

In terms of performance, the comparison of global and distributed/semi-distributed approaches in modeling is a problem that has been strongly developed for a long time (Wending 1992; Ague et al. 2014). Results from such model assessments provide a relatively complex picture as some authors clearly state the benefits of using the distributed/semi-distributed approach (Michaud & Soroshian 1994; Krysanova et al. 1999; Boyle et al. 2001) while other studies (Diermanse 1999; Kokkonen & Jakeman 2001) present opposite results. Some studies also show that the global approach gives better results in small watersheds, while the distributed/semi-distributed approach performs better in the case of large watersheds. These findings suggest that distributed/semi-distributed models are particularly applicable for complex watersheds due to their physical heterogeneity (Tegegne et al. 2017). Due to the complex nature of our study area, this study adopted a semi-distributed modeling approach considering that a previous study in the area using the lump conceptual modeling approach produced poor results (Sighomnou 2004). One of the complexities of the study area is the existence of two rainy seasons in the region.

In recent decades, several distributed/semi-distributed hydrological models have been developed to simulate the hydrological processes of watersheds and predict flows. Typical examples of distributed hydrological models include TOPMODEL (Topography-based Hydrological Model) (Beven & Kirkby 1979), SHE (European Hydrological System) (Abbott et al. 1986), Soil and Water Assessment Tool (SWAT) (Arnold et al. 1998), MGB-IPH (Large Basin Hydrological Model) (Collischonn & Tucci 2001), etc. Since the others only allow an approximate characterization of the physical environment of the watershed through the use of data and parameters in a point-grid network (Cao et al. 2006; Wang et al. 2012), SWAT appears to be the most efficient model in the wide range of applications.

The SWAT model has been widely used around the world for hydrological simulations of basins with different environmental conditions (climate, topography, geology, soils and vegetation) (Von Stackelberg et al. 2007; Pereira et al. 2010; Andrade et al. 2013) with satisfactory results obtained. For SWAT to become a universal hydrological model, more studies are needed in basins with climatic regimes and soils typical of equatorial conditions, such as that of the So'o basin in southern Cameroon.

The objectives of this study were to (1) evaluate the capacity of the SWAT model to simulate flows in a complex equatorial river basin with two rainy season regimes and (2) use the model to simulate future flows in the basin under different climate change scenarios. One of the biggest challenges to hydrological modeling in basins in the absence of sufficient flow gauging stations and the ability of the SWAT to simulate flows in this poorly gauged basin will be of great importance for socioeconomic development considering the number of construction projects including dams and bridges ongoing in the basin. The results obtained could help to improve the management of water resources in this basin and the region.

The spatial data required for this study are of the following three types: a Digital Elevation Model (DEM), land use map and soil map.

The FAO world digital soil map downloaded from the site http://www.fao.org/geonetwork/srv/en/metadata.show?id=14116 was used as soil data in this study. The soil classification used is based on the FAO classification system and was customized as required by the SWAT model (Figure 2). Three types of soils characteristic of the forest zone, namely the Orthic Ferralsols (Fo), have been identified in the basin (Fo1-ab–481, Fo26-ab–490 and Fo33-2ab–491). Types Fo26-ab–490 and Fo33-2ab–491 are the most widely represented with 46.7 and 53.2% of the whole basin.

Apart from topographic and soil information which remain unchanged on the human time scale, the simulation of flows from the SWAT model requires other dynamic information such as land use. Three Landsat satellite images were used to produce land cover maps. These are the Landsat Thematic Mapper (TM) images of 2005, and Landsat 8 (OLI) of 2015 and 2020. The maps produced from the images of 2005 and 2015 served as a reference in the QGIS software, to predict the land cover in 2035 and 2065. The 2020 map was used to validate the one simulated for the same year from the two previous ones (2005 and 2015 maps) since any prediction from a model requires calibration and validation of the said model. Similarly, the reliability of a model's outputs depends on the validation results of the model in question.

Rainfall data used to study the current period were obtained from the Climate Research Unit (CRU) of the University of East Anglia in the United Kingdom. The data are available from 1901, via the site https://climexp.knmi.nl/selectfield_obs2.cgi?id=2833fad3fef1bedc6761d5cba64775f0/ in NetCDF format, on a monthly time step and at a spatial resolution of 0.25°×0.25°. The flow series used is that of ORE-BVET (Environmental Research Observatory/Tropical Experimental Watersheds).

The following meteorological data are for modeling flows in SWAT at daily time step: maximum and minimum temperatures (°C), precipitation (mm/day), relative humidity (%), average wind speed (m/s) and solar radiation (in W/m²). Among all these variables, only rainfall was observed at the selected stations (Nsimi, Ebolowa and Sangmelima) (Figure 1). Precipitation data were obtained from the meteorological service of the Ministry of Transport in Cameroon. The other variables were uploaded to these same stations on the SWAT website (https://globalweather.tamu.edu/). These are data from the National Centers for Environmental Prediction (NCEP). It has been shown that these data could constitute a good alternative for modeling flows in ungauged regions (Dile et al. 2013; Fuka et al. 2013; Nkiaka et al. 2017). All the meteorological data used were collected over the period 1998–2008.

The analysis of average rainfall and flow was carried out using Pettitt (Pettitt 1979) and modified Mann–Kendall (Hirsch & Slack 1984; Araghi et al. 2014) statistical tests at the 95% significance level. Following the application of autocorrelation (from the calculation of the R statistic) and seasonality (from the employment of the correlogram) tests to the rainfall and flow series used, it turned out that there is a seasonality in the latter. It is why we chose to use the modified Mann–Kendall test of Hirsch & Slack (1984) to the detriment of the classic Mann–Kendall test and other modified Mann–Kendall tests.

If the null hypothesis is rejected, an estimate of the break date is given by the instant defining the maximum in absolute value of the variable Ut, N.

There is no significant trend in the series analyzed when the calculated p-value is above the chosen significance level. More details on the modified version of the Mann–Kendall test can be found in the relevant literature (Hirsch & Slack 1984).

The response of discharges from the So'o watershed to climate change and anthropization was assessed over two future periods (2022–2060 and 2061–2100). The interval 1999–2008 served as the reference/historical period. Rainfall data of the station used are without gaps over this interval. The 2005 land cover map was used to simulate runoff for the reference period (1999–2008). Those simulated for the years 2035 and 2065 were, respectively, used to simulate basin flows in the near (2022–2060) and distant (2061–2100) future.

Cellular automata (CA)–Markov is the procedure used for land cover prediction in this work. It combines Markov chains (quantity), multi-criteria evaluation (location) and filtering. This procedure is described as CA (Halmy et al. 2015). Markovian chains analyze two images of land cover at different dates and produce two transition matrices (probability and affected area in pixels for persistence and transition), and a set of conditional probability images. They make it possible to calculate a future state from the known present state, based on the observation of past evolutions and their probability. This makes this method one of the best for modeling the temporal and spatial dimensions of land use patterns (Halmy et al. 2015; Yang et al. 2019).

In this study, two regional climate models (RCMs) [RCA4 and CCCma] from the CORDEX project that have proven to be effective in simulating precipitation and temperature in Africa (Gadissa et al. 2018; Dibaba et al. 2019) were retained. However, despite their reliability and the degree of confidence that can be granted to them, the outputs of the models sometimes present considerable biases, which require corrections before using them to study the impact of climate change. For each of the RCMs, data from two scenarios (RCP4.5 and RCP8.5) were collected. The first and second scenarios are, respectively, representative of high greenhouse gas emissions and moderate emissions. The other meteorological variables (solar radiation, relative humidity and wind speed) considered for the historical period have been taken over for the two future periods without making any changes, given that their modifications have no significant impact on the modeling result (Gadissa et al. 2018).

Climate Model Data for Hydrological Modeling (CMhyd) software (Rathjens et al. 2016) obtained from https://swat.tamu.edu/software/ was used to correct for precipitation and temperature biases. Teutschbein & Seibert (2012) provided a comprehensive review of bias correction techniques based on this tool. According to the authors, all the correction techniques improved the simulations of precipitation and temperature. However, they noted differences between the correction methods. Based on the proximity between the corrected datasets and the observed datasets, distribution mapping (DM) was considered to be the best correction method, both for temperature and precipitation. According to the authors, DM uses a transfer function to adjust the cumulative distribution of the corrected data to that of the observed data, which makes the results significantly better. Zhang et al. (2018) compared five bias correction methods using the CMhyd tool. They demonstrated that DM was the most efficient correction method for studying the impact of climate change on the flow dynamics of two rivers in the northern basin of Lake Erie (Canada). Based on these results, DM was retained for the precipitation and temperature corrections of the model outputs used in this study.

SWAT is a physically based semi-distributed hydrological model, designed and developed by researchers at the USDA (United States Department of Agriculture) (Arnold et al. 1998). The physical aspect of the model makes it possible to reproduce the processes that take place in the environment, using different sets of equations (Neitsch et al. 2005; Arnold et al. 2012). This model is continuous over time and is designed to run simulations over long periods (Payraudeau 2002). The SWAT model analyzes the watershed as a whole by subdividing it into sub-watersheds containing homogeneous portions called HRUs. Each HRU is characterized by unique land use, soil type and topography. SWAT provides access to the different water balance components at the HRU scale for each time step (daily, monthly and annual) over the simulation period (Neitsch et al. 2005).

The validity of a hydrological model was checked by comparing the model simulated (Q_sim) (Q_cal) and observed (Q_obs) flows through subjective and quantitative criteria. Initially, a good match between the observed and simulated flow hydrographs will attest to good calibration. In the second step, we used three of the most widely used criteria for the validation of hydrological models, correlation coefficient (r), the Nash index (NSE) and the bias (Akoko et al. 2020).

The average annual and seasonal flows of the So'o decreased over the periods 1998–1999 and 2017–2018 (Figure 3). These decreases are all non-significant according to the Pettitt test. The mean summer flows that experienced the largest decrease are the only ones to have recorded a significant decrease according to the Mann–Kendall test (Figure 3). The smallest decrease in runoff was observed during the spring season.

The impact of rainfall on the evolution of the So'o flows seems quite clear at annual and seasonal time scales. We note for the two compared variables the same evolution during the period 1998–1999 to 2018–2019, which makes it possible to consider a possible role of the decrease in rainfall on runoff. It should be noted, however, that the magnitudes of the decreases in precipitation in the different seasons do not always correspond to those of the runoff in the same seasons. The summer dry season flows, for example, recorded the largest drop, yet it was for the autumn rainfall that the largest drop was observed. This could be linked to the fact that in the equatorial region, the precipitation of certain seasons has an impact on the flows of other seasons, most often those which follow them (Liénou et al. 2008; Ebodé et al. 2022).

To identify the parameters having a major influence on the outputs of the model, a sensitivity analysis was carried out from the daily flows observed. Initially, four parameters relating globally to runoff and groundwater were identified as being the most sensitive in calibration (Table 1).

This work confirms the ability of the SWAT model to reproduce flows satisfactorily in the equatorial region. Similar results were obtained elsewhere. For example, Arnold & Allen (1999) have used observed data from three watersheds, ranging in size from 122 to 246 km², to successfully validate flows simulated from SWAT. Other authors (Arnold et al. 1999) have also successfully evaluated the ability of the model to reproduce flows in the Gulf of Texas over basin sizes between 2,253 and 304,260 km².

To highlight the evolution of the future climate, the averages of the corrected precipitation and temperature series for each of the two RCMs (CCCma and RCA4), for each of the two emission scenarios (RCP4.5 and RCP8.5) and for each future period (2022–260 and 2061–2100) were compared with those of the historical period (1999–2008).

In their study of the Finchaa watershed in Ethiopia, Dibaba et al. (2020) highlighted a decrease in rainfall from the RCA4 model in the near (2021–2050) and distant (2051–2080) future. This study, however, notes a decrease in precipitation for the same model in the near future only (2022–2060). An increase is projected by the model in the distant future (2061–2100).

The CCCma model predicts a decrease in precipitation in the SRB for both periods and under the two scenarios considered (Table 2). For the first period, the decreases are projected to be, respectively, –31 and –30.8% under the RCP4.5 and RCP8.5 scenarios. For the second, respective decreases of –27.9 and –18% are forecast under the RCP4.5 and RCP8.5 scenarios (Table 2). This decrease in precipitation will mainly result from that of the summer and autumn precipitation (Figure 6). Under the RCP8.5 scenario, the 1940s is the only decade for which the deficit seems more severe compared to the others, three decades present an apparent deficit under the RCP4.5 scenario including the 2020s, 2040s and 2070s (Figure 7).

The two models (RCA4 and CCCma) project an increase in maximum and minimum temperatures, regardless of the period and the scenario considered (Figure 6 and Table 3). In general, this temperature rise will gradually increase over the decades, and only peak towards the end of the century (Figure 7). The RCP8.5 scenario also appears to be the one for which the increase is greater, especially after the 1960s (Table 3 and Figure 7). Under the RCP8.5 scenario, the average maximum temperature of the 2100 decade exceeds 35°, regardless of the model considered. However, it does not exceed 32° under the RCP4.5 scenario (Figure 7). The observation is practically the same for minimum temperatures. Those predicted by the two models under the RCP8.5 scenario are generally about 2° higher than those predicted under the RCP4.5 scenario during the last two decades of the century. Seasonally, both models predict a larger increase in winter and spring minimum temperatures. For maximum temperatures, on the other hand, the expected increase is almost regular throughout the year (Figure 6).

As is the case in this work, several studies dealing with the impact of climate change on resources (Kingston & Taylor 2010; Basheer et al. 2015) have already noted in the hydrological units studied a gradual increase in future maximum and minimum temperatures under the different scenarios retained.

Three main modes of land use having a direct link with the flow have been identified in the SRB. These are the forest, urban areas and water bodies. The land cover forecasts for the years 2035 and 2065 were certainly made based on the changes noted between 2005 and 2015, but the year 2005 is considered as the reference year from which the probable future land cover changes are assessed.

The predictions of climate models (precipitation and temperature) were integrated into the SWAT model to get an idea of possible future changes in discharges in the SRB.

With regard firstly to the RCA4 model, the results obtained show that a change in precipitation in line with projections will cause an increase in discharges, whatever the period and the scenario taken into account. In general, this increase will be greater during the dry seasons (winter and summer), and the decades at the end of the century will be the wettest (Figures 6 and 7). The changes (increases) noted during the second period (2061–2100) are consistent with those observed for rainfall under the two scenarios (RCP4.5 and RCP8.5). We could also see an effective impact of climate change on flows during this period and under these two scenarios. The climate forecasts of this model do not, however, explain the increase noted during the first period (2022–2060) under the same scenarios. This increase is indeed concomitant with a decrease in precipitation and an increase in temperature. The analysis of the impact of anthropogenic changes and its effects which will be done later in this part of the article will certainly allow us to see more clearly on this subject.

Concerning the CCCma model, a change in precipitation in line with the forecasts of the latter will cause a decrease in the runoff, except under the RCP8.5 scenario during the second period (2061–2100), where we note an increase compared to the historical period of about +4% (Table 2). This decrease will mainly concern the autumn months, and the 2020s, 2040s and 2070s will be the most affected (Figures 5 and 6). The impact of climate change is also visible on the flows simulated from the outputs of the model, for almost all periods and scenarios. It is only for the second period under the RCP8.5 scenario that we observed a different evolution between precipitation (decrease) and flow (increase) (Table 2). This again gives rise to questions about the possible impact of anthropogenic change and its effects on future flows of the So'o.

To ensure better management of water resources, the impact of climate change water resources has been addressed in several studies around the world (Notter et al. 2013; Wagena et al. 2016; Danvi et al. 2018; Duku et al. 2018). In these different studies, the outputs of the climate models were integrated into the SWAT model to simulate future flows and get an idea of their possible evolution. As is the case in this study, some of the studies also predict identical trends in the evolution of future precipitation and runoff (Beyene et al. 2010; Basheer et al. 2015; Dibaba et al. 2020).

The explanations for certain changes in the future flows of the So'o watershed could not be justified by projected changes in precipitation and temperature. In this case, the projected changes in simulated flows for the first period (2022–2060) from the RCA4 model under the two scenarios, and the simulated flow for the second period (2061–2100) from the outputs of the CCCma model under the RCP8.5 scenario. In these different scenarios, an increase in discharges has been noted, while rainfall decreases and temperatures increase, which should rather cause a decrease in discharges, due to a reduced supply of the groundwater and increased evapotranspiration. This raises questions about the evolution of surface conditions, knowing that an increase in impermeable spaces, for example, increases runoff and therefore flows.

The objectives of this study were to evaluate the capacity of the SWAT model to simulate flows in a complex equatorial river basin with two rainy season regimes and to use the model to simulate future flows in the basin under different climate change scenarios. Between 1998–1999 and 2018–2019, the SRB experienced a decrease in discharges caused by a decrease in precipitation. Model simulation using outputs from the CCCma model indicates that flows in this river will decrease more, except under the RCP8.5 scenario during the second period (2061–2100), where flows are projected to increase compared to the historical period by approximately +4%. This decrease will be observed mainly in the autumn months, and the decades 2020, 2040 and 2070 will be the most affected. A change in precipitation according to the projections of the RCA4 model will on the other hand cause an increase in discharges by more than +50%, whatever the period and the scenario. This increase will be greater during the dry seasons (winter and summer), and the decades at the end of the century (2080–2100) will be the wettest. For some cases, an increase in flows was noted despite a drop in rainfall, particularly in the case of flows simulated for the first period from the outputs of the CCCma model. This seems to be a consequence of the increase in urban areas. This is confirmed by the increase in runoff observed during this period.

V.B.E., J.G.D. and J.J.B. conceptualized the study and performed the methodology; V.B.E. did software analysis and formal analysis; investigated the study; did data curation; V.B.E., E.N. and J.G.D. wrote and prepared the original draft; V.B.E., B.N.N. and E.N. wrote, reviewed and edited the article; project administration by V.B.E. All authors have read and agreed to the published version of the manuscript.

This research received no external funding.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.