The Sissili-Kulpawn Basin (hereafter SKB) consists of the Sissili Basin (hereafter SB) and the Kulpawn Basin (hereafter KB), which are adjacent sub-basins of the White Volta Basin located in West Africa. The SKB has an estimated area of 23,576 km² with major parts in Ghana and some parts in Burkina Faso. The SB is located between 10.28°N–12.0°N and 2.58°W–1.8°W (Figure 1), covering a 12,800 km² area, and a core research site of the West Africa Science Service Center on Climate Change and Adapted Land Use (WASCAL) (Arnault et al., 2016a; Graf et al., 2021). The topography is mostly flat, ranging from 300 m to 400 m in the Northern Burkina side of the basin and from 200 to 300 m in the Ghana side of the Basin (Figure 1). The presence of a protected wildlife area in the central parts of the basin (Nazinga Game Ranch, where no farming activities occur) ensures land use and land cover changes in the SB are less pronounced (Bliefernicht et al., 2018). Rainfall within the SB is unimodal, with a pronounced wet season ranging between May and September. Annual total rainfall is about 1,200 mm. Runoff from rainfall drains into the Sissili River, with the principal discharge station at Wiasi (Figure 1). The KB lies between 9.6°N–11.1°N and 2.6°W–1.2°W almost entirely within Ghana, covering about 10,541 km² (Figure 1). The topography is uniformly flat over the Basin, with elevations from 200 to 250 m. The main land use activity is subsistence crop production such as guinea corn, millet, cotton, groundnut, sorghum, and animal husbandry. The average rainfall for May–June is 114 mm and about 202 mm for August–September, based on the Climate Research Unit gridded Time Series (CRU TS) datasets over the 1981–2020 period. Runoff generated from rainfall flows into the Kulpawn River which is measured at Yagaba.

The MODIS dataset in WRF-Hydro was replaced by the ESA CCI LC data (Defourny et al., 2017) for the synthetic land cover change numerical experiment. The ESA CCI reference LC data was designed to meet LC desires expressed by the climate modeling and to avoid false change detection between LC classes that are semantically close, hence its adoption for this study. Moreover, it has a relatively high spatial resolution compared to the default LC dataset in WRF-Hydro (MODIS). To validate the WRF-Hydro simulated discharge, daily observed discharge at Yagaba (−1.283, 10.23) on the Kulpawn River was provided by the Volta Basin Authority (VBA) at Ouagadougou, spanning from 1st January 2010 to 31st December 2016. The WRF-Hydro simulated temperature and rainfall were validated using the Climate Hazards Group Infrared Temperature with Stations (CHIRTS) dataset and the Integrated Multi-satellitE Retrievals for GPM (IMERG), respectively. Though the performance of satellite data products varies from Basin to Basin, the IMERG is the second-best performing rainfall dataset over Africa at daily timescale, hence its adoption for this study (Mekonnen et al., 2023). CHIRTS-daily dataset also performed similar to or better than ERA5 and ERA5-Land at eight stations in Africa (Parsons et al., 2022). Since ERA5 was used to force the WRF-Hydro, it was incorrect to use it to validate the model’s output. In that sense, the CHIRTS was used to validate the simulated temperature over the SKB. Due to the scarcity of land-atmosphere exchange energy flux datasets, the globally available energy flux dataset from FLUXCOM is used to validate the WRF-Hydro simulate fluxes, including the net radiation flux (R^net), sensible heat flux (H^sensible), and latent heat flux (H^latent). The data described above are detailed in Table 1.

The WRF regional climate model (version 4.4) and the WRF-Hydro hydrological module (version 5.2) were coupled to describe in more complex detail the regional water cycle of the SKB. The setup follows the one used in Arnault et al. (2023) and consists of two domains (Figure 1); an outer and an inner domain. The outer domain is at 50 km resolution, covering an area of 6,000 km x 4,000 km, including western and parts of central Africa, with 50 pressure levels up to 10 mbar. The inner domain encompasses the SKB at 10 km horizontal resolution, covering 800 km x 800 km, with 50 pressure levels up to 10 mbar. The lateral boundaries and initial conditions of the outer domain are forced with ERA5 reanalysis (Hersbach et al., 2020) atmospheric fields (geopotential height, meridional and zonal winds, water vapor, air pressure, temperature) at 0.25° resolution and six-hourly time steps. To ensure numerical stability, the atmospheric equations of motion of the inner and outer domains are resolved at 60 s and 180 s, respectively. The selected physics parameterization options for the inner and outer domains in Table 2, including terrestrial hydrology, radiation, cloud microphysics, turbulence, and cumulus convection, are based on performance in reproducing well the daily basin-averaged rainfall of the SKB as explored in various configurations. The impact of different parameterization combinations is not the subject of this study. A comprehensive overview of northern Sub-Saharan Africa is given by Laux et al. (2021).

The WRF-Hydro hydrological module makes possible the inclusion of a lateral flow in the inner domain (Gochis et al., 2021). Using input elevation and hydrological data from version 2 of HydroSHEDS database (Lehner et al., 2008) with version 5.2 of WRF-Hydro GIS Pre-processing Tool, the inner domain is coupled with a subgrid of 1 km resolution (see Figure 1B). The minimal stream number for defining the channel network in Figure 1B is 25. The coupling procedure includes the aggregation/disaggregation of the surface water and soil moisture variables between the inner domain grid and the subgrid with so-called disaggregation factors. The disaggregation factor is defined by the ratio of the fine grid variable’s value to its corresponding coarse grid value. For every time step, the disaggregation factors are used to disaggregate the Noah-MP soil moisture, soil, and surface water variables onto the subgrid, routed in river channels, overland, and in the subsurface based on diffusive wave formulations (Gochis et al., 2021), and then reaggregated again to the Noah-MP grid. The fully coupled formulation described above made the comprehensive description of a basin-scale water cycle possible.

The SKB is prone to perennial flooding (Gross and Pennink, 2018), with the possibility of the rivers going beyond their banks. The default WRF-Hydro does not account for such overbank flow, which often results in unrealistically high peaks in the simulated discharge, compared to much smoother discharge peaks in the observed. To circumvent this issue, Arnault et al. (2023) modified the WRF-Hydro source code by including an overbank flow parameter that allows for a two-way flow of water between the river and land and applied it successfully to the Nzoia river basin in tropical East Africa. In this work, we now apply the calibrated overbank flow option proposed by Arnault et al. (2023) to obtain an improved simulation of the observed discharge. A detailed explanation of how simulated discharge is improved and numerical balance achieved with the overbank flow option is outlined in Arnault et al. (2023). The source code can be downloaded at https://doi.org/10.6084/m9.figshare.21063982. A four-year calibration period starting from 1st January 2010 to 31st December 2013 was used for calibrating the WRF-Hydro simulated discharge. The calibration was achieved using the WRF-Hydro enhanced with an overbank flow parameter H_thresh (in meters) to reduce the unrealistically high discharge peaks by allowing water originating from an upstream channel pixel to flow towards the land surface once the water head in the channel pixels exceeds H_thresh (Arnault et al., 2023). The three most sensitive parameters including the percolation parameter S, the overbank flow parameter H_thresh, and the river roughness Manning coefficients were manually tuned in the calibration. The H_thresh smooths the discharge peaks but also removes much water from the channels, which is corrected by decreasing the S to partially seal the soil column bottom and force water to exfiltrate back to the surface. Reducing the river Manning coefficients tends to reduce water accumulation in the streams so that H_thresh is less often reached thus further modulating the overbank flow effect. The approach to this calibration strategy is detailed in Arnault et al. (2023). To validate the WRF-Hydro simulated discharge, the results were compared with observed discharge at Yagaba Station on the Kulpawn River using evaluation metrics outlined in section 2.5.

Four goodness-of-fit metrics are employed to show various aspects of the performance of the WRF-Hydro simulated outputs (Equations 1–5). To evaluate the mean rainfall, temperature, water and energy fluxes in space (inner domain) and over time (simulated period), it suffices to compute the percentage bias (PBIAS) between the simulated and observational products. The PBIAS is used to assess if the model is overestimating or understanding the observational data, with 0 indicating no bias between the observed and modeled data, and positive and negatives indicating overestimations and underestimations, respectively. Besides the PBIAS, the coefficient of determination (R²), the Nash–Sutcliffe efficiency (NSE), and Kling–Gupta efficiency (KGE) metrics are used to numerically compare the simulated and observed rainfall and streamflow timeseries. The R² is used to assess the goodness of fit of the observed rainfall/streamflow to simulated rainfall/streamflow with values of R² close to 1 indicating a perfect fit. The NSE is the traditional metric used in hydrology to summarize model performance and takes values from − ∞ to 1 with NSE ≥ 0.5 indicating satisfactory performance. With NSE ≥ 0.7, the model can be considered to have a very good fit (Nash and Sutcliffe, 1970). The KGE provides a diagnostically interesting decomposition of the NSE which facilitates the analysis of the relative importance of different components (bias, correlation, and variability) in hydrological modeling (Gupta et al., 2009). Like the NSE, the KGE ranges from − ∞ to 1 with a value close to 1 indicating a more accurate model. (1) R 2 = ∑ i = 1 n Q o b s − Q ¯ o b s × Q s i m − Q ¯ s i m ∑ i = 1 n Q o b s − Q ¯ o b s × ∑ i = 1 n Q o b s − Q ¯ s i m 2 2 (2) PBIAS = ∑ i = 1 n Q o b s − Q s i m ∑ i = 1 n Q o b s × 100 (3) N S E = 1 − ∑ i = 1 n Q o b s − Q s i m 2 ∑ i = 1 n Q o b s − Q ¯ o b s 2 (4) K G E = 1 − r − 1 2 + α − 1 2 + β − 1 2

Where β a bias term, α a measure of the flow variability error, and r is the linear correlation between observations and simulations. (5) K G E = 1 − r − 1 2 + σ s i m σ o b s − 1 2 + μ s i m μ o b s − 1 2

where σ_sim is the standard deviation in simulations, σ_obs is the standard deviation in observations, μ_obs is the observation mean Q ¯ o b s , and μ_sim is the simulation mean Q ¯ s i m . The Q s i m and Q o b s terms also refers to the simulate and observed parameters, respectively.

To assess the potential impacts of land cover changes from afforestation and deforestation on regional water and energy budgets, a reference simulation from 1st January 2010 to 31st December 2016 is generated using the updated and calibrated WRF-Hydro (see section 2.4). The result of the reference experiment is validated with observational data products presented in section 2.2, and the results are discussed in sections 3.1–3.8. As indicated by De Noblet-Ducoudré et al. (2012) the evaluation is necessary to assess whether the model is good for numerical landcover change experiments. For the reference experiment, the 2010 ESA LC map is used (see Figure 2A). Figures 2D,E is a resampling of the modified ESA LC (Figures 2B,C) to the WRF-Hydro grid, based on the land cover change experiments. The cropland experiment (see Figure 2B) represents a replacement of closed shrubland (43.49%) within the SKB area by cropland such that the total cropland area is 71.7% (Table 3). This is a deforestation scenario as it turns to increase albedo and reduce the leaf area index (LAI) compared to the reference (compare Figures 3A,B,D,E). The closed shrubland experiment (see Figure 2C) represents a replacement of the cropland (28%) by closed shrubland such that the total closed shrubland area is 72% (Table 3). This scenario is an afforestation scenario in which albedo is reduced and LAI increased (compare Figures 3A,C,D,F). Aside from the dominant land cover types mentioned above, all other land cover types within the inner domain remain unchanged. The land cover change numerical experiment was carried out over 7 years from 1st January 2010 to 31st December 2016. The seven-year period is considered sufficient to obtain a robust modeled climate signal given the idealized nature of the experiment and the fact that ESA LC maps did not change significantly over a twenty-eight-year period (1992–2019). The seven-year experiment was conducted for each of the two land cover change scenarios and the differential results with respect to the reference scenario discussed in sections 3.6–3.8. The land cover changes numerical experiments give a signal of the realistic climatic impact that could be associated with afforestation and deforestation scenarios and the extent to which climatic impacts depend on vegetation cover change. Figure 2 shows the spatial representation of modifications made for each land cover change numerical experiment. Table 3 gives the total land cover area in km² and the percentage of land cover type within the SKB, SB, and KB. Table 4 represents the total land cover area over the SKB for each land cover type in each of the two numerical experiments.

The different land cover change experiments result in different mean albedo and LAI values compared to the reference WRF-Hydro simulation as seen in Figures 3A,D. The WRF-Hydro simulated mean albedo values for cropland, and closed shrubland scenarios over the SKB are 0.25, and 0.17 respectively, compared to a mean reference value of 0.20. This indicates the cropland scenario results in an increase in mean albedo over the SKB by +0.05, while the closed shrubland scenario results in a decrease in the mean albedo over the basin by −0.03, compared to the reference scenario (Figure 3A). The fact that the basin average albedo value for the cropland scenario is higher than that of the closed shrubland scenario is consistent with literature as replacing closed shrubland with cropland is a deforestation scenario that leads to an increase in albedo. Croplands have higher albedo values (0.18–0.25) than closed shrublands (0.16–0.18) which is depicted by our results. The LAI and albedo value changes are in phase opposition, with the LAI increasing (decreasing) as the albedo decreases (increases). The LAI for the cropland, and closed shrubland experiments over the SKB are 1.55 m²m⁻², and 2.03 m²m⁻², respectively, the reference value being 1.83 m²m⁻². The change of the LAI under the cropland, and closed shrubland are, respectively, −0.28 m²m⁻², and 0.20 m²m⁻², indicating a reduction in LAI for the cropland scenario and an increase in LAI for the closed shrubland scenario compared to the reference scenario (Figure 3D). From the analysis of the albedo and LAI, our model represents well these two biophysical properties of land-atmosphere interactions necessary for simulating the effects of land cover changes on the energy and water budgets.

The results of the skill metrics for simulated precipitation and discharge are shown in Figures 4A,B, respectively. The default setup values are indicated in grey and are lower than the corresponding calibrated setup values. For precipitation, the R², NSE, KGE, and PBIAS for Kulpawn Basin are 0.56, 0.31, 0.64, and − 2.8%, respectively. For daily discharge, these performance values are 0.53, −0.34, 0.2, and − 52.4%, respectively. Thus, the default uncalibrated setup yields relatively weak discharge metrics compared to precipitation, emphasizing the need for calibrating the surface hydrology component. This is evident in the unrealistically high peaks in the simulated discharge in grey (see Figure 4B).

Figure 5 shows the validation of the calibrated WRF-Hydro simulated rainfall and temperature with IMERG and CHIRTS observational products. The WRF-Hydro simulated mean temperature and precipitation over the SKB are 27.3°C and 2.7 mm day⁻¹, compared to 29.2°C and 2.6 mm day⁻¹ from CHIRTS and IMERG observational products, respectively. The bias is thus −1.7°C and 3.5% showing an underestimation of temperature and overestimation of precipitation. The calibrated WRF-Hydro simulated rainfall shows biases compared to the observational dataset (IMERG) within the range of −25 to +25% (Figure 5C) with underestimations mainly in the northern parts of the inner domain within the upper parts of Sissili in Burkina Faso. Lower overestimations within −10 to 10% are mainly constrained to the southern parts covering the entire KB and the lower parts of SB. The relatively lower bias of simulated rainfall in the KB could explain the relatively good performance in the discharge evaluation metrics at Yagaba on the Kulpawn River, with the R², NSE, KGE, and PBIAS of 0.53, 0.47, 0.69, and − 15%, respectively. Yet, the moderate performance in simulated discharge could be attributed to the quality of the simulated precipitation (see Figure 5C), a well-known limitation of the coupled atmospheric-hydrological modeling approach (Senatore et al., 2015). The moderate NSE metrics fall within the range of published precipitation and discharge performance across the globe: 0.46 for (Arnault et al., 2016a), 0.27 for (Kerandi et al., 2018), −1.89–0.61 for (Rummler et al., 2019), 0.27 for (Li et al., 2020), 0.30 for (Arnault et al., 2023). More importantly, the calibrated setup improves both rainfall and discharge over the SKB compared to the default setup (see metrics in Figures 4A,B). Consequently, the rainfall-discharge results obtained over the calibration period are considered realistic and the WRF-Hydro with two-way river-land flow option is adopted to investigate the climate response to land cover changes in this work. The biases in precipitation show that the fully coupled WRF-Hydro simulated temperature and precipitation biases are minimal such that the model is skilled in providing climate information on daily timescales. In West Africa, rainfall and temperature are the most basic information provided by national meteorological organizations to the public. With the capabilities of the fully coupled WRF/WRF-Hydro shown in this work, national meteorological agencies can use the WRF model to complement or improve the traditional approach to weather forecasting. That said, the WRF-Hydro simulations are computationally expensive, and strong computational resources and skills are required to run simulations. The employees of the national meteorological agencies must be skilled to simulate, analyze the results, and make reports available to the public as climate information or weather forecasts. The traditional way of forecasting and climate information provision could significantly benefit from the fully coupled WRF/WRF-Hydro.

The percentage bias of the WRF-Hydro simulated R^net, H^sensible, and H^latent fluxes in the inner domain ranges between −25 to 10%, −10 to 50%, and − 50 to 10%, respectively, (see Figures 6C,F,I), over the 2013–2016 period. Over the SKB, the average percentage bias of R^net, H^sensible, and H^latent is 11, 31% and − 4% indicating an overestimation of R^net and H^sensible but an underestimation of the H^latent flux. In terms of numerical values, the mean R^net averaged over SKB is 135 Wm⁻² compared to 121 Wm⁻² from the FLUXCOM observation product. The mean simulated H^sensible is 70 Wm⁻² compared to 53 Wm⁻² from the FLUXCOM observation product. Finally, the mean H^latent flux is 65 Wm⁻² compared to 69 Wm⁻² over the 2010–2013 period. The overestimation of H^sensible and underestimation of H^latent suggests that the modeled land compartment is too dry.

The climatology of WRF-Hydro simulated rainfall over the SKB is close to IMERG observational product, with mostly underestimations, except for a slight overestimation of June–July–August–September (JJAS) rainfall (Figure 7A). Thus, the WRF-Hydro with riverbank overflow option improves the simulated rainfall but is a little wetter than that observed in JJAS over the SKB. The impact of the JJAS simulated rainfall on discharge is resolved by the overbank flow option in the updated WRF-Hydro such that the discharge in JJAS stays very close to the observed discharge (Figure 7B), with only a slight overestimation in September. The WRF-Hydro simulated temperature also underestimates the CHIRTS observational product but stays very close in the April–May–June–July (AMJJ) period (Figure 7C). In terms of R^net, the WRF-Hydro simulation mimics well the observational product FLUXCOM for January–February–March–April (JFMA) and June–July–August (JJA) with the most overestimations at May–June (MJ) and September–October–November–December (SOND) periods. The simulated H^sensible overestimates the observational product for all months, while the H^latent underestimates the FLUXCOM data for nearly all months of the year except December (Figure 7D). From the foregoing, the simulated rainfall (P), temperature (T), discharge, R^net, H^sensible, and H^latent show that the WRF-Hydro with overbank flow extension as used in this study mimics reasonably the hydroclimatic situation of the SKB and can be considered for providing climate service for the region.

The average changes in the R^net, H^sensible, H^latent over the SKB for the cropland scenario are obtained from Figures 8A–C and evaluated as −11 Wm⁻², −9 Wm⁻², and − 2 Wm⁻², respectively. For the closed shrubland, the changes in R^net, H^sensible, and H^latent over the SKB is evaluated from Figures 8D–F as +7 Wm⁻², +5 Wm⁻², and + 2 Wm⁻², respectively. From the results above, on the one hand, the cropland scenario decreases the R^net, H^sensible, and H^latent fluxes compared to the closed shrubland scenario which results in a net increased energy flux. As mentioned previously, the decrease in R^net, H^sensible, and H^latent fluxes for the cropland is related to the increase in the albedo (see Figure 3B) and a decrease in LAI (see Figure 3E) which reduces R^net, H^sensible, and H^latent fluxes (see Figures 8A–C). In contrast, the decreased albedo (see Figure 3C) and increased LAI (see Figure 3F) for the closed shrubland scenario increase absorption of incoming radiation but decreases outgoing radiation to space, thus resulting in surplus energy near the surface which increases R^net, H^sensible, and H^latent fluxes (see Figures 8D–F).

The change in precipitation and temperature averaged over SKB between the cropland and the reference scenario estimated is −0.1 mm day⁻¹ and + 0.1°C respectively. For the closed shrubland scenario, the change in precipitation and temperature over the SKB is +0.1 mm day⁻¹ and -0.1°C. Accordingly, there is an increase in temperature and a decrease in precipitation for the cropland but a decrease in temperature and an increase in precipitation for the closed shrubland scenario. For the closed shrubland scenario, the surface temperature increase through enhanced sensible heating is offset by an enhanced evaporative cooling from the soil and plant surface (see Figures 10E,F). Consequently, the temperature for the closed shrubland scenario decreases as can be seen in Figure 9D. The evaporative cooling also enhances atmospheric moisture and consequently increases precipitation for the closed shrubland scenario (see Figure 9B). Conversely, the cropland scenario behaves as a deforestation scenario where there is an increase in albedo but a decrease in LAI. As a result, net radiation (see Figure 8A) and heat fluxes (see Figures 8B,C) are decreased and less energy is available for evaporative cooling through soil and plant evaporation, consequently increasing the temperature (Figure 9C). The reduced evaporation affects the amount of moisture in the atmosphere to precipitate back as rainfall, hence the observed reduction of rainfall for the cropland scenario (see Figure 6A). It is noted that non-radiative forcing is the dominant factor affecting temperature in the tropics (Zhang et al., 2022), such that afforestation usually results in a decrease in temperature (Betts, 2000; Li et al., 2015). Lawrence and Vandecar (2015) using satellite towers, and ground-based observations have shown that tropical deforestation results in warmer, drier conditions at the local scale. The closed shrubland scenario, which decreases albedo and increases LAI similar to an afforestation scenario, tends to decrease temperature and increase precipitation, which confirms the strong correlation between land surface and climate variables discussed by Hardwick et al. (2015). Conversely, the cropland scenario which tends to increase albedo and decrease LAI is consistent with a deforestation scenario, resulting in an increase in temperature and a decrease in the precipitation. From the foregoing, we conclude that our numerical experiment can provide a sound prediction of the impacts of afforestation and deforestation scenarios on regional temperature and precipitation. The results show that afforestation or deforestation could modulate the local climate by decreasing or increasing temperature and rainfall. The synthetic numerical experiments conducted in this work are extreme cases of deforestation and afforestation with positive and negative impacts (see Tables 3, 4 for the percentage area afforested/deforested). In that sense, the community can benefit from a controlled afforestation scenario that prevents the adverse effects of flooding associated with the extreme afforestation scenario. Conversely, farmers or the SKB project could benefit from controlled cropland expansion that does not significantly affect rainfall. Land managers could benefit from the threshold values of deforestation and afforestation used in this work to guide land management in the basin. This practice will ensure the climate-land-water-energy-food balance in the basin with no adverse effects.

For the cropland scenario, the ΔE^soil, ΔE^plant, ΔR^surface_, ΔR^ground over the SKB are estimated from Figures 10A–D as −0.05 mm day⁻¹, −0.01 mm day⁻¹, −0.03 mm day⁻¹, and − 0.0004 mm day⁻¹, respectively. In terms of the closed shrubland scenario, ΔE^soil, ΔE^plant, ΔR^surface_, ΔR^ground are estimated as +0.03 mm day⁻¹, +0.03 mm day⁻¹, +0.03 mm day⁻¹, and + 0.00006 mm day⁻¹ (see Figures 10E,F). We recall a slightly higher net positive change in R^net, H^sensible, and H^latent for the closed shrubland scenario compared to the cropland (see Figure 8). These surface energy flux changes enhance soil evaporation and plant transpiration in the closed shrubland scenario (see Figures 10E,F) compared to the cropland scenario (see Figures 10A,B). The enhancement of the soil evaporation and plant transpiration increases atmospheric moisture to precipitate back as rainfall leading to higher surface runoff in the closed shrubland scenario (see Figure 10G). Conversely, in the cropland scenario, the decrease in soil evaporation and plant transpiration reduces moisture in the atmosphere to precipitate back as rainfall consequently reducing the surface runoff (see Figure 10C). In both scenarios, however, there is an insignificant change in underground runoff over the SKB. The climatological water flow change of the closed shrubland afforestation scenario (see Figure 11A) shows an average increase in precipitation of +6%, evapotranspiration of +3%, surface runoff of +27%, and underground runoff of +16% (see Table 5). In the cropland scenario, the simulated annual water flow changes of precipitation, evapotranspiration, surface runoff, and underground runoff over SKB all decreased by −5, −3%, and − 9%, respectively, (Table 5) relative to the reference scenario. In general, our numerical land cover change experiment realistically mimics the regional water cycle response to changes in biophysical properties.