The Contribution of Transpiration to Precipitation Over African Watersheds

Abstract

The redistribution of biological (transpiration) and non‐biological (interception loss, soil evaporation) fluxes of terrestrial evaporation via atmospheric circulation and precipitation is an important Earth system process. In vegetated ecosystems, transpiration dominates terrestrial evaporation and is thought to be crucial for regional moisture recycling and ecosystem functioning. However, the spatial and temporal variability in the dependency of precipitation on transpiration remains understudied, particularly in sparsely sampled regions like Africa. Here, we investigate how biological and non‐biological sources of evaporation in Africa contribute to rainfall over the major watersheds in the continent. Our study is based on simulated atmospheric moisture trajectories derived from the Lagrangian model FLEXPART, driven by 1° resolution reanalysis data over 1981–2016. Using daily satellite‐based fractions of transpiration over terrestrial evaporation, we isolate the contribution of vegetation to monthly rainfall. Furthermore, we highlight two watersheds (Congo and Senegal) for which we explore intra‐ and interannual variability of different precipitation sources, and where we find contrasting patterns of vegetation‐sourced precipitation within and between years. Overall, our results show that almost 50% of the annual rainfall in Africa originates from transpiration, although the variability between watersheds is large (5%–68%). We conclude that, considering the current and projected patterns of land use change in Africa, a better understanding of the implications for continental‐scale water availability is needed.

Key Points

• We show the contribution of transpiration to precipitation over 25 African watersheds based on satellite data and Lagrangian modeling

• Transpiration contributes to more than 50% of the precipitation over the African continent, although the variability among watersheds is large

• The Congo and Senegal show contrasting trends in precipitation, seemingly associated with different contributions of transpiration and oceanic evaporation

1. Introduction

Ecosystems around the world are coping with increasing water stress due to climate change and anthropogenic disturbance. Global warming intensifies the hydrological cycle through increasing evaporation rates and precipitation intensity (Ficklin et al., 2019; Westra et al., 2014). Furthermore, rising atmospheric carbon concentrations enhance plant productivity (leading to “global greening,” see Piao et al., 2020), affecting hydrology at local to global scales (Lu et al., 2016; Zeng et al., 2018). Meanwhile, anthropogenic disturbance through direct extraction of surface‐ and groundwater leads to a global overexploitation of water resources (Wada et al., 2010), while land cover change has affected a third of the global land area, mostly in the direction of reducing vegetation cover (Winkler et al., 2021). While studies indicate that the effects of land cover change on the continental water cycle may be buffered by shallow groundwater (Zipper et al., 2019), land cover change has led to a general loss in the capacity of ecosystems to capture and retain water (Gerten et al., 2005), and a reduced flux of terrestrial evaporation (E, often referred to as “evapotranspiration”) (Sterling et al., 2013). E is pivotal for the distribution of water on our planet, being a key contributor to precipitation (P) over land (Eltahir, 1998; Tuinenburg et al., 2020; Van Der Ent et al., 2010). It is estimated that globally, 70% of terrestrial E rains back over the continents (Tuinenburg et al., 2020), which amounts to 40% of the total terrestrial P (Van Der Ent et al., 2010). The importance of terrestrial moisture recycling varies both in time and space, with some regions being highly dependent on this “self‐supplied” moisture (Holgate et al., 2020; Keys et al., 2014). In Amazonia, for instance, the contribution of local E is important to maintain forest stability and buffer against droughts. About 20% to 45% of P over Amazonia originates from the region itself (Burde et al., 2006; Staal et al., 2018; Trenberth, 1999). Similarly, in the Congo watershed, up to half of P derives from the watershed itself (Sorí et al., 2017; Tuinenburg et al., 2020). It should be noted that moisture recycling metrics are scale‐dependent (see Section 3.3) and therefore not directly comparable. Estimates of seasonal variability are, however, highly uncertain, ranging from 25% to 83% of wet season P originating from the Congo region itself (Dyer et al., 2017; Worden et al., 2021). In temperate regions and drylands, the sign and magnitude of moisture recycling has been less investigated than for tropical forests (Wierik et al., 2021), although some studies suggest the contribution of local E is equally crucial for P supply in these regions (Miralles et al., 2016; Nieto et al., 2006; Savenije, 1995; Yu et al., 2017). Beyond self‐supplied P, studies show moisture recycling cascades spanning over continents (Staal et al., 2018; Zemp et al., 2014), complex networks of atmospheric moisture exchange between regions (Keune & Miralles, 2019; Keys et al., 2017), and spatial hydrological connectivity through atmospheric pathways that supply moisture to densely inhabited regions (de Vrese et al., 2016; Keys et al., 2018). The large variation in the estimations following precipitation recycling studies is partly explained by the methods used, ranging from isotope ratios (Gat, 1996; Zhao et al., 2019), to various models, such as bulk recycling models (Burde et al., 2006; Pokam et al., 2012) or Eulerian and Lagrangian transport models (e.g., Tuinenburg et al., 2020; Van Der Ent et al., 2010), often employing different input data (Tuinenburg & Staal, 2020).

The increasing availability and resolution of satellite observations, land surface models (LSMs), isotope data, and field measurements enables the disaggregation of E into its different components (Stoy et al., 2019). Broadly speaking, the terrestrial E flux can be partitioned into the direct evaporation from the soil surface, intercepted rainfall, and open water bodies (hereafter collectively referred to as “E _d ”), and the biological flux of transpiration (hereafter “E _t”) (Miralles et al., 2020). In general, vegetation cover regulates E _d and E _t in two major ways. First, passively, by changing the water and energy balance at the land surface, that is, through effective rainfall partitioning (Crockford & Richardson, 2000), changing the radiation partitioning (He et al., 2014; Otto et al., 2011), rainfall interception (Horton, 1919), changes in leaf area index (LAI) (Wang et al., 2014), and modifying the water retention capacity of the soil (D’Odorico et al., 2007). Second, actively: through biophysically regulated stomatal conductance and root growth, vegetation modulates the magnitude and timing of E _t (Jarvis et al., 1976). A multitude of methods has been developed to disaggregate E, but estimations of the global contribution of E _t to terrestrial E are highly uncertain, with reported values ranging between 24% and 90% due to the variety of methods applied (Wei et al., 2017). In general, isotopic analyses appear to provide estimates that are at the high end of that range: for example, an analysis of isotopic data of rivers and lakes around the world led to an estimated contribution of 80%–90% of E _t to E (Jasechko et al., 2013). Yet, Good et al. (2015) estimated E _t to E to be only 64% based on global isotopic mass balance of the atmosphere and ocean. In contrast, LSM tend to provide lower estimates: for example, Wang‐Erlandsson et al. (2014) analyzed global E partitioning using a hydrological LSM and found a 59% contribution of E _t to E. Remote sensing based approaches, on the other hand, vary widely in regards to the E _t to E ratio depending on the model used (Miralles et al., 2016). By combining various estimates for different vegetation types, Wei et al. (2017) presented a composite approach that suggested an average 57% contribution of global E _t to E. Regional differences in this ratio are expected to occur among different ecosystems and climates (Green et al., 2017; Miralles et al., 2016; Pranindita et al., 2021).

Although vegetation is considered a particularly important contributor to precipitable moisture (Meier et al., 2021; O’Connor et al., 2021), only few moisture recycling studies explicitly differentiate between biological and non‐biological evaporation fluxes. Existing studies at the global level used partitioned E estimates from a LSM (Wang‐Erlandsson et al., 2014), coupled to an atmospheric moisture tracking model, to estimate spatially distributed moisture recycling metrics (van der Ent et al., 2014), or applied a “green earth” and “desert earth” scenario to estimate the contribution of vegetation to global P (Keys et al., 2016). The former study estimated that 56% of global E _t returns to the land surface as P (van der Ent et al., 2014); the latter concluded that global vegetation cover regulates 20% of the total P over land. Based on several studies, a review paper from Schlesinger and Jasechko (2014) estimated that E _t accounts for 39% of terrestrial P through its contribution to precipitable atmospheric moisture. At the regional level, the direct contribution of vegetation to P through E _t is particularly poorly understood. Over Africa in particular, moisture recycling studies are sparse (Pokam et al., 2012), despite the current and future challenges posed by land use change, climate change and water scarcity in the region (Ahmadalipour et al., 2019; Haile et al., 2020; Held et al., 2005; Herrmann et al., 2020).

Considering the limited understanding of the dynamics of “vegetation‐sourced P,” as well as sparse moisture recycling assessments over Africa, this study has two objectives. First, to identify the contribution of vegetation‐sourced P patterns over Africa, defined here as terrestrial P originating from E _t worldwide. We analyze the source regions of monthly P (over 1981–2016) for each of the 25 major African watersheds (sink regions) at 1° resolution. We use watersheds because they are a functional unit that is often used in hydrology. We distinguish between oceanic and terrestrial moisture sources, the latter disaggregated into terrestrial moisture sources from within the watershed (i.e., local moisture) and outside the watershed (i.e., remote moisture). Furthermore, we differentiate between E _t and E _d contributions to P. Second, to present and describe the development of a data set following the analysis above; this data set is available online to spur further research on specific drivers and dynamics of moisture recycling over Africa (see Data Availability Statement). We use two watersheds (Congo and Senegal) to showcase some of the research possibilities offered by the available data. As such, we aim to contribute to enhanced understanding of the role of vegetation in the regional‐to‐continental water cycle and provide moisture recycling metrics for individual African watersheds.

2. Methodology

2.1. Data Generation: Identify Rainfall Source Regions Using FLEXPART

Our study is based on atmospheric moisture trajectories simulated with the Lagrangian model FLEXPART v9.01 (Stohl et al., 2005) and driven with ERA‐Interim reanalysis data (e.g., Drumond et al., 2014; Nieto et al., 2019). The spatial resolution of the ERA‐Interim forcing employed here is 1°, with 61 vertical levels between the land surface and 0.1 hPa (Dee et al., 2011). FLEXPART simulations are performed globally over 1981–2016 (36 years), initialized with a homogeneously distributed parcel density and 2 million parcels—on a 1° grid, this comprises approximately 30 parcels over each grid cell at each time step. While 3‐hourly forecasts are used to supplement the 6‐hourly reanalysis data to improve the accuracy of the simulated trajectory, only 6‐hourly reanalysis time steps are used for the analysis. Simulations of FLEXPART driven with ERA‐Interim reanalysis have been used in various studies (see, e.g., Table 1) and facilitate the identification of moisture fluxes with reasonable accuracy (Keune et al., 2022). While the use of higher‐resolution forcing data sets may be desirable and promising (Hoffmann et al., 2019), moisture tracking studies that employ ERA‐Interim as a forcing remain state‐of‐the‐art for large‐scale applications and have shown high agreements to other reanalyses, such as MERRA (Bosilovich et al., 2011) over the Western Sahel (Keys et al., 2014). Here, we evaluate moisture trajectories from FLEXPART over the 25 major African watersheds from HydroSHEDS (Lehner & Grill, 2013). Major watersheds were chosen to ensure that enough parcels can be tracked. For example, over the Senegal basin — one of the smaller basins that roughly comprises almost 5 · 10⁵ km² or ∼40 1° × 1° grid cells—around ∼4,800 parcels are evaluated each day. To estimate the origins of precipitation, we applied the moisture‐tracking framework by (Keune et al., 2022) to the outputs from the FLEXPART simulations that are global in extent. The framework comprises three steps that are executed for each watershed over the full climatology: diagnosis, attribution, and bias‐correction. A short description of each step is presented below; for details, see Keune et al. (2022).

Table 1.

Non‐Exhaustive Overview of Moisture Recycling Studies Over the African Continent

Reference	Area	Main findings	Methods/models	Input data source
Savenije (1995)	Sahel	Moisture recycling in the Sahel account for 90% of the rainfall Irrigation during the dry season does not necessarily lead to increased precipitation due to low atmospheric vapor	Simple linear analytical model based on Budyko (1974)	Rainfall gauges and runoff observations (field‐based)
Gong and Eltahir (1996)	West Africa (5°N–15°N, 10°W–15°E)	Precipitation in West Africa derives for 23%, 27%, and 17% on evaporation from the Atlantic Ocean, West Africa and Central Africa, respectively Moisture supply from Central Africa is limited by the westerly monsoonal circulation	Mass balance based on Budyko (1974)	ERA‐40, 2.5°, daily, 1985–1995
Mohamed et al. (2005)	Nile Basin	Mean annual local moisture recycling of 11%	Regional Atmospheric Climate MOdel (RACMO), applying Budyko framework, run for 35.96°N–12°S and 10°E−54.44°E, 0.44°	RCM initialized with ERA‐40, 2.5°, 1995–2000Vegetation cover from GLCC GlobalLand Coverage CharacteristicsDischarge data (gauged measurements at 11 stations)Precipitation from ground stations; GPCC, 1°, monthly; FEWS, 0.1°
Los et al. (2006)	Sahel	Vegetation‐rainfall feedback explains up to 30% of the rainfall variability between 15° and 20°N	Statistical Vegetation Index Simulation (SVIS)	NDVI from AVHRR, 0.5°, monthly, 1982–1999Precipitation and temperature from CRU, 0.5°, monthly, 1901–2000
Van der Ent et al. (2010)	African continent	West‐Africa is a major (precipitation) sink of continental evaporation Indian Ocean provides moisture to East and Central Africa, which in its turn provides moisture to West Africa Total precipitation from continental (terrestrial) origin over Africa = 49%	Water accounting model	ERA‐Interim, 1.5°, 6 hourly, 1998–2008
Pokam et al. (2012)	Equatorial Central Africa (5°N–5°S and 12.5°E–30°E)	Seasonal variability in moisture convergence determined by African Easterly Jet location and Atlantic Ocean Mean annual recycling ratio of 0.38 Low annual cycle/variability of recycling ratio	2D bulk recycling model based on (Burde et al. (2006)	National Center for Environmental Prediction/National Center for Atmospheric Research (NCEP/NCAR) Reanalysis, 1.9° and 2.5°, (monthly., at 1.9° × 1.9° and 2.5° × 2.5°)
Keys et al. (2014)	Western Sahel	Variation in Sahelian E most important to explain variance of precipitation 50.1% of the precipitation in the “core precipitation shed” derives from the land surface	Eulerian Water Accounting Model 2 Layers (WAM‐2Layers, v2.3.01),	ERA‐Interim, 1.5° and MERRA, 1.0° × 1.25°, 1979–2013
Salih et al. (2015)	Sahelian Sudan	ITCZ is the main source of moisture During July and August, the ITCZ brings in half of the precipitation Despite being relatively dry, trade winds from the Arabian Peninsula are responsible for almost 30% of the precipitation	FLEXPART v6.2	ERA‐Interim, 1.5° and 2°, 3/6 hourly, 1998–2008
Miralles et al. (2016)	East and West Sudanian Savanna, Serengeti, Kalahari Desert	East Sudanian Savanna receives almost all P from terrestrial E Recycling ratio Kalahari Desert during growing season 34% Local recycling ratios increase during dry years (intensification of dry conditions)	FLEXPART v9.0	ERA‐Interim, Evaporation from GLEAMand OAFLUXVegetation Optical Depth (VOD), all data regridded to 0.25° with monthly time steps
Oguntunde et al. (2016)	Niger Basin	Local recycling between June and September amounts 21%	Regional Climate Model (RegCM3) over 1981–2000	ERA‐InterimPrecipitation and temperature from CRU TS 2.1 (0.5°, monthly) River discharge from gauging stations
Dyer et al. (2017)	Congo Basin (10°S–5°N and 15°E−30°E)	Local E contribution varies between 24% and 38% throughout the year	Water tagging capability in NCAR Community Earth System Model (CESM v1.2), run at 1.9° × 2.5° (30 vertical levels)	CRU, GPCP, ERA‐Interim, and GPCC precipitation data sets
Sorí et al. (2017)	Congo Basin	The Congo Basin provides over 50% of the total atmospheric moisture for precipitation High oceanic E events are not directly linked to increased P over the basin Dry/wet years are associated with lower/higher local moisture contribution	FLEXPART v9.0	Precipitation from CRU TS v3.23, 0.5°Runoff from GRDCEvaporation from GLEAM v2 and OAFLUXERA‐Interim, 1°, 6 hourly
Yu et al. (2017)	Sahel	Oceanic forcing (i.e., SST) dominates precipitation variability (22% annual mean), except during post‐monsoon period (SON) Annual mean terrestrial forcing of 8%, with 18% in SON	A multivariate, lagged Generalized Equilibrium Feedback Assessment (GEFA)	Various land and ocean variables, see Yu et al. (2017) supplementary material for full overview
Keys and Wang‐Erlandsson (2018)	Niger	41% of annual precipitation derives from the ocean 9% derives from Niger itself, mostly from the South of the country	WAM‐2layers	ERA‐Interim, 1.5°, 3/6 hourly
Wang‐Erlandsson et al. (2018)	Congo Basin	Half of the streamflow in the Congo river is depending on terrestrial moisture recycling	STEAM and WAM‐2Layers	Meteorological data with ERA‐Interim, 1.5°, 3/6 hourly, 1995–2014Land Cover data with Ramankutty potential land‐cover, IGBP (MCD1C1), MIRCA2000 v1.1Precipitation from MSWEP v1, 0.25° regridded to 1.5°Runoff data from GRDC, 0.5° regridded to 1.5°
Notaro et al. (2019)	Sahel, Greater Horn of Africa, WAM‐region, Congo basin	Land–atmosphere feedbacks are most prominent in semi‐arid regions Forcing impact of soil moisture and vegetation is comparable of oceanic forcing for the whole region	A multivariate Stepwise Generalized Equilibrium Feedback Assessment (SGEFA)	Various land surface variables, see Table 1 in Notaro et al. (2019) for full overview
Tuinenburg et al. (2020)	Congo basin	Evaporation recycling of 60% Precipitation recycling of 47%	UTrack, analysis at 0.5°, monthly	ERA5, 0.25°, hourly, 2008–2017
Worden et al. (2021)	Congo basin	83% of free tropospheric moisture comes from terrestrial E in February (first rain peak) During the second rain peak (September–October), only 31% derives from terrestrial E	Statistical model based on satellite observations of deuterium content in water vapor	Deuterium content from NASA Tropospheric Emission Spectrometer; SIF from GOME‐2 V26 740 nm data productsPrecipitation from TRMM, 0.25°.E from MODIS and ERA5, 0.25°, daily

First, daily fluxes of E and P over each 1° pixel are diagnosed by tracking changes in air parcel properties from the Lagrangian simulation using all global two‐step trajectories (Keune et al., 2022). Here, the atmospheric moisture balance is used to estimate E and P from diagnosed changes in the specific humidity of air parcels, subjected to specific process‐based detection criteria to filter for each process (diagnosis). Compared with observational reference data sets from MSWEP, OAFLUX, and GLEAM (Beck et al., 2019; Miralles et al., 2011; L. Yu & Weller, 2007), see description below, this evaluation of all parcels produces a data set representing accuracy and reliability that will be used in the third step (bias‐correction). Second, in the attribution step, all air parcels arriving or residing over an individual watershed are selected and traced back for 15 days. Here, only precipitating parcels (i.e., parcels with relative humidity >80% losing moisture, following the convection parameterization from Emanuel (1991), are selected and their 15‐day trajectories are evaluated to determine source–sink relationships of moisture. Note, however, that while we consider 15‐day trajectories, the actual residence time of moisture in the atmosphere is not prescribed but determined through a discounting procedure at the parcel level (see, e.g., Sodemann et al., 2008; Läderach & Sodemann, 2016). Along each individual 15‐day trajectory, all moisture gains occurring in the maximum atmospheric boundary layer between two time steps are considered to represent source locations associated with E; we focus on the primary surface sources of precipitation only and neglect any secondary sources (such as re‐evaporation in the atmosphere). Along each trajectory and between each identified surface source location and the P event, rain en route is accounted for through linear discounting (Sodemann et al., 2008) and ensures mass balance closure along the trajectories that are used to establish the source regions. The latter procedure assumes that the atmospheric boundary layer is well mixed, and it renders the prescription of the residence time unnecessary. Finally, in the third step, a daily bias‐correction against observational data of E and P is performed using three observational reference datasets for each 1° grid cell (for source region E) and at the watershed level (for sink region P). As only primary source–sink relationships associated with surface fluxes E and sink fluxes P are considered, both fluxes can be bias‐corrected using observational data (Keune et al., 2022). The absolute E flux in each source region pixel (as determined from in the diagnosis step above) is used to bias correct E in the source region (Keune et al., 2022); therefore, the relative contribution of each source pixel contribution over the absolute E flux in the same pixel is calculated, and bias‐corrected over the ocean with OAFLUX (Yu & Weller, 2007), and over land using GLEAM v3.5a (Martens et al., 2017). P in the sink region (i.e., the watershed) is bias‐corrected using MSWEP v.2.2 (Beck et al., 2017, 2019). We apply MSWEP because it outperforms ERA‐Interim P estimates over West and East Africa (Sahlu et al., 2017; Satgé et al., 2020). Moreover, Satgé et al. (2020) find that MWSEP best represents daily precipitation dynamics over West Africa compared to other precipitation data sets, such as CHIRPS (Funk et al., 2015). Considering the use of daily P fluxes in the analytical framework described above, we consider MSWEP most suited, despite limited capacity of earlier MSWEP versions (2.1) to represent regional hydroclimatic extremes and uncertainties over Africa (Awange et al., 2019). MSWEP is also used as input for GLEAM, which we expect to guarantee some water balance consistency.

Additionally, to unravel the contribution of E _t to P, we disaggregate the source contribution E in each grid cell into E _d and E _t based on daily E _d /E and E _t /E ratios from GLEAM, respectively. GLEAM estimates have been validated using eddy‐covariance towers and in situ soil moisture measurements across the world (Martens et al., 2017) and have been used in several studies to benchmark other models (e.g., Dong et al., 2022; Lian et al., 2020). Chen et al. (2022) find that GLEAM may overestimate E _t , especially during the wet season and in low vegetation cover regions. Nonetheless, Wei et al. (2017) show that GLEAM partitioning of evaporation and transpiration fluxes compares well to ensemble means and other global products. In short, the setup employed here tracks all precipitating parcels backwards and identifies all biological and non‐biological moisture sources within the atmospheric boundary layer to establish the spatial source–sink relationships of moisture (Keune et al., 2022). The P data (in mm) is aggregated at the monthly level and is available from January 1981 up until December 2016 (see Data Availability Statement).

2.2. Data Analysis: Spatial and Temporal Variability of Moisture Source Regions

For each watershed, we analyze the mean monthly and mean annual contribution of five different moisture sources: oceanic E, local E _t and E _d , and remote E _t and E _d (Figure 1). We apply a global land surface mask to distinguish between ocean and terrestrial sources, use daily E _t and E _d fractions to distinguish between E _t and E _d fluxes (aggregated to monthly values), and employ watersheds from HydroSHEDS to distinguish between local and remote sources. Subsequently, based on individual watershed E _t contributions to P, we estimate the mean annual contribution of E _t to P (%) over the continent (E _t,c ) as:

$$E_{t,c}=\frac{{\sum }_{i=1}^n{\bar{P}}_{t,i}}{{\sum }_{i=1}^n{\bar{P}}_i}\times 100$$

where ${\bar{P}}_i$ is the mean annual rainfall (km³ year⁻¹) in watershed i (n = 25). We use flows (km³) instead of fluxes (mm) to enable aggregation of the vegetation‐sourced P contributions from all watersheds. Meanwhile, ${\bar{P}}_{t,i}$ is the ${\bar{P}}_i$ originating from E _t (km³ yr⁻¹). Both ${\bar{P}}_i$ and ${\bar{P}}_{t,i}$ represent the mean values over the full climatology (1981–2016).

Figure 1.

Sources of mean annual P over major African watersheds averaged over 1981–2016. (a) Shades of green represent vegetation‐sourced P over the watershed; shades of yellow represent precipitation originated from E _d . Blue represents the ocean‐originated precipitation. The size of each pie‐chart represents the magnitude of mean annual precipitation. (b) Contribution (%) of each watershed to the total continental E _t ‐sourced precipitation. Figure S1 in Supporting Information S1 shows the corresponding names of the watersheds.

For each watershed, we distinguish between annual, wet season and dry season contributions of E _t . We identify wet (W _i) and dry (D _i) seasons for each watershed individually, based on mean monthly P for a particular month j being higher (wet season) or lower (dry season) than the mean monthly P for the whole record (${\bar{P}}_i$), similar to the method to distinguish wet and dry years described by (Keys et al., 2018):

$$W_{i}if{\bar{P}}_{j,i}>{\bar{P}}_i$$

and

$$D_{i}if{\bar{P}}_{j,i}<{\bar{P}}_i,$$

where ${\bar{P}}_{i,j}$ represents the mean precipitation of a specific month j in watershed i, and ${\bar{P}}_i$ the mean monthly precipitation in watershed i (mm year⁻¹). Both variables represent the mean values over the full climatology.

3. Results and Discussion

3.1. Variability of Rainfall Sources Over Major African Watersheds

Our results indicate that in most major African watersheds, a significant part of P is vegetation‐sourced (Figure 1a). Overall, almost 50% of P over the African continent derives from transpiration worldwide (E _t,c ). This aligns with the 56% of global terrestrial P estimated to originate from E _t in van der Ent et al. (2014), but it is higher than the global estimate from Schlesinger and Jasechko (2014) of 39%. However, the latter is based on a data set that hardly sampled African regions. Figure 1b shows how much of the continental vegetation‐sourced precipitation can be attributed to individual African watersheds and underlines the disproportionate contribution of the Congo basin.

The variability in annual E _t dependency between watersheds is large: between 5% and 68% of annual P is vegetation‐sourced (Figure 2a). In most coastal watersheds along the Mediterranean and Southern Africa coast less than 20% of the annual P originates from E _t , as the majority derives from oceanic evaporation (Figure 1). In other regions, such as the tropical Congo basin, around 68% of the mean annual P derives from E _t , the majority (52%) originating from the watershed itself (Figure 1). In general, continental watersheds, such as Lake Chad and Zambezi, are more dependent on terrestrial E compared to oceanic E, with the majority of P originating from E _t .

Figure 2.

Seasonal variability of watersheds dependency on vegetation‐sourced precipitation. (a) Variability in mean annual, wet season and dry season contribution of E _t to P between watersheds (note that numbers represent min, max, first quartile, median and third quartile). (b) Relative contribution of local E _t (left column) and remote E _t (right column) for the wet season (upper row) and dry season (lower row).

During the dry season, the average dependency of P on E _t is slightly lower (Figure 2a). However, for some watersheds, E _t appears to be particularly important during the dry season. For example, in Senegal, the relative contribution of E _t to P increases from 36% during the wet season to 62% during the dry season (see Figure 2 and Figure S2 in Supporting Information S1)—indicating that vegetation is important to maintain the dry season P over the watershed. Overall, we find different seasonal dependency‐regimes of vegetation‐sourced P (Table 2) that may be influenced by land–atmosphere feedbacks coupled to seasonal climatic events (Green et al., 2017). For example, Yu et al. (2017) find that vegetation‐rainfall linkages are exceptionally strong in the post‐monsoon period over the Sahel. Furthermore, Pokam et al. (2012) show that seasonal cycles of moisture convergence, driven by the location of the African Easterly Jet, increase recycling ratio's in the dry season. For some regions, there is a constant contribution of E _t throughout the year (i.e., the difference of E _t contribution between the wet‐ and dry season <5%), whereas in other regions, E _t is particularly important in the dry (e.g., Senegal) or wet season (e.g., Namibia/Swakop). Figure 2 distinguishes between wet‐ and dry season E _t originating from local E _t and remote E _t . Some watersheds (e.g., Congo) show a strong dependency on local E _t throughout the year and rely little on other terrestrial source regions for both the dry and the wet season P. Other watersheds such as Zambezi and Lake Chad, which are in the vicinity of the Congo watershed, are dependent on remote E _t, which indicates that vegetation in the Congo is an important source for P in other watersheds (Van Der Ent et al., 2010).

Table 2.

List of Watersheds (n = 12) Where More Than 30% of the Mean Annual P Is Vegetation‐Sourced (Mean Annual E _t Dependence in %)

Watershed	Mean annual E t dependency (%)	Interannual variability vegetation‐sourced P (COV)	Seasonal dependence on E t (constant/wet/dry)	Mean annual P (mm year−1)
Congo	>50%	Low	Constant	1478
Nile	>50%	Low	Constant	614
Zambezi	>50%	High (dry season)	Wet season	853
Lake Chad	>50%	Moderate (dry season)	Dry season, terrestrial	370
Rift Valley	>50%	Low	Constant	757
Africa South Interior/Okavango	>50%	High (dry season)	Constant	476
Senegal	>30%	Moderate (dry season)	Dry season	575
Namibia/Swakop	>40%	High	Wet season	82
Orange	>40%	High (dry season)	Wet season	296
Africa East Central Coast	>40%	High (dry season)	Wet season (local)	923
Niger	>30%	Moderate (dry season)	Dry season (remote)	710
Gulf of Guinea	>30%	Low	Wet season	1814

Note. Interannual variability of vegetation‐sourced P is defined here based on three value classes for the coefficient of variation: <20% (low); 20%–40% (moderate); and high (>40%). We distinguish between annual mean, wet and dry season. Seasonal dependence on E _t is based on the difference in relative contribution of E _t between the wet and dry season. To classify the watersheds based on the seasonal dependence, we use a threshold of 5% relative difference between the wet and dry seasons. It is considered constant if the difference between mean dry and wet season contribution <5%. The threshold value aims to standardize seasonal differences and is arbitrarily chosen. The watersheds in italic (Congo and Senegal) will be highlighted further.

3.2. Temporal Variability in the Congo and Senegal Watersheds

Table 2 highlights watersheds with a high (>30%) mean annual dependency on E _t for precipitation. The watersheds are described based on mean annual E _t dependence (%), interannual variability in E _t contribution using the coefficient of variation (COV, Figure S3 in Supporting Information S1), seasonal dependence on E _t (Figure 4 and Figure S2 in Supporting Information S1), and climatology (mean annual P in mm year⁻¹). Based on these criteria, we select two watersheds with a different hydro‐climatology and describe their annual and seasonal source region characteristics in detail below: the Congo and Senegal watershed. Note that the spatially explicit source regions depicted in Figures 3 and 5 are defined based on “significant contributions” of grid cells that contribute (on average) more than 5 mm year⁻¹ of E to P over the sink region (similar to the method applied by Keys et al. (2014).

Figure 4.

(a) Annual P (mm) over the Congo basin between 1981 and 2016. (b and c) Annual contribution of E _t to P in mm and fractional contribution to total rainfall, respectively. (d and e) Annual contribution of E _d in mm and fractional contribution to total rainfall, respectively. Full lines represents local sources, dashed line represents remote sources. (f and g) Annual contribution of oceanic E to P in mm and fractional contribution to total rainfall, respectively. Note the different scaling of the x‐axes corresponding to local (left y‐axis) and remote (right y‐axis) sources of E _t and E _d contributions.

Figure 3.

(a) Mean moisture source regions for the Congo watershed (black delineation). All moisture source regions <5 mm are considered insignificant contributors and are not visible on this map (in white); (b) Mean moisture source regions during the dry season (May–August); (c) Mean moisture source regions during the wet season (September–April); (d) Intra‐annual absolute contributions of different moisture sources (mm, left axis) and P (mm, right axis, dashed line); (e) Intra‐annual relative contributions of different moisture sources (%); all averaged over the full climatology.

Figure 5.

(a) Mean moisture source regions for the Senegal basin (black delineation). All moisture source regions <5 mm are considered insignificant contributors and are not visible on this map (in white); (b) Mean moisture source regions during the dry season; (c) Mean moisture source regions during the wet season; (d) Intra‐annual absolute contributions of different moisture sources (mm, left axis) and P (mm, right axis, dashed line); (e) Intra‐annual relative contributions of different moisture sources (%); all averaged over the full climatology.

3.2.1. Congo Watershed

The Congo watershed is one of the wettest watersheds in Africa (Table 1). It is located around the equator and covered by 2 million km² of tropical forest (Hansen et al., 2008). Our results indicate that most of the annual P is sourced from E from the watershed itself (68%) (Figure 3a). This is similar to the estimate from Sorí et al. (2017), who applied FLEXPART and ERA‐Interim over 1980–2010 and found that, throughout the year, monthly P constitutes for 60% of E from the Congo itself. However, estimations from Tuinenburg et al. (2020) using the Lagrangian moisture tracking model UTrack (forced with ERA5) are substantially lower (47% of P derived from local E). Similarly, Dyer et al. (2017) apply a Eulerian moisture tracking scheme within a NCAR Community Earth System Model (CESM) over the Congo basin, but estimate that local E contribution varies between 24% and 38% throughout the year.

Only 11% of the mean annual P derives from the ocean, both the Indian Ocean on the east side of the continent, and the South Atlantic along the west coast. Despite the stable relative contributions of different moisture sources throughout the year (Figure 3e), there is a clear dip in the amount of P between May and August (i.e., the dry season), which corresponds to a reduction in the contribution of local E _t to rainfall (Figure 3d). Two distinct rainy seasons emerge in the boreal fall and spring (Figure 3d), which are associated with the passing of the intertropical convergence zone (Notaro et al., 2019). Overall, the relative contribution of local E _t to P is persistently high throughout the year (around 50%; Figure 3e). Further, even during the dry season, the absolute contribution of local vegetation‐sourced P is dominant (Figure 3d), and the relative contribution is slightly increased (Figure 3e)—thus indicating that local vegetation maintains P over the Congo basin during the dry season even when oceanic moisture convergence is lower. In line with findings from Worden et al. (2021), increased advection of ocean air in the September–November rainy season slightly reduces the relative contribution from land E.

Addressing the changes in P and the contribution of different moisture sources between years (Figure 4), we observe three remarkable patterns. First, we find a decreasing trend in annual P (Figure 4a) as derived from MSWEP that seems to correspond primarily to decreasing oceanic moisture contributions (Figure 4f), in contrast to observed global trends of increasing ocean contributions (Findell et al., 2019), and as noted in previous studies (Gimeno, Nieto, & Sorí, 2020). Sorí et al. (2022) confirm a drying trend, particularly in the northern part of the Congo basin. A similar and persistent drying trend over the Congo was found by Cook et al. (2020) using multiple P data sets. They found reduced P over the Congo from 1979 onwards and explained this trend with shifting thermal lows that weaken moisture convergence in the (boreal) spring and fall; in the boreal summer, enhanced drying is associated with increasing warming over the Sahara (Cook et al., 2020) and an increasing length of the dry season (Jiang et al., 2019). Others have attributed the trends of observed drying over equatorial Africa (including the Congo) to the Atlantic Multidecadal Oscillation changing sea surface temperatures (SSTs) over the North Atlantic, warming over the Indian Ocean, and tropical biomass burning further causing increasing carbonaceous aerosol concentrations in the atmosphere (Diem et al., 2014). As such, drivers of observed P reductions remain debated. Second, our results indicate a slight increase in absolute and relative contributions of local E _t (Figures 4b and 4x) along with a reduction in E _d contributions to P (Figure 4d). The increasing relative contributions of local E _t is evident during both wet and dry seasons (Figure S4 in Supporting Information S1). This implies that contributions of local E _t to P increase, and vegetation‐sourced E becomes more important for P over the Congo. This unexpected increase of vegetation‐sourced P in light of an increasing drying trend may be explained by biological responses at the plant level. Asefi‐Najafabady and Saatchi (2013) found that progressive drying had no particular effect on tropical tree cover, suggesting that forests have adapted gradually to increasing drought conditions. This could imply that forests accessing deeper groundwater can maintain transpiration when precipitation is low (Pranindita et al., 2021). Increased E _t may also result from increased plant density with enhanced productivity. For example, Zeng et al. (2018) find that in response to rising atmospheric carbon concentrations (i.e., CO₂ fertilization), increasing LAI around the globe has led to increased evaporation and moisture recycling, particularly in wet regions. Nonetheless, the increasing E _t contributions are affected by the uncertainties in the GLEAM estimates of E _t and E _d , which are based on observed vegetation fractional cover and do not explicitly account for potential reductions in stomatal conductance in response to rising atmospheric CO_2. Consequently, the contribution of local E for P trends over the Congo remains subject to debate (Dyer et al., 2017; Sorí et al., 2017; Worden et al., 2021). At last, we observe that around the year 2000, there is a reduction of both relative and absolute contribution of remote E _d and E _t (Figures 4b–4e) which may be explained by altered hydroclimatic conditions outside the Congo basin (Cook et al., 2020).

3.2.2. Senegal Watershed

The Senegal watershed is situated along the coast of West‐Africa and receives a mean of 575 mm rainfall per year. On average, 39% of P is vegetation‐sourced, with the majority coming from outside the basin (24% from remote E _t ; 15% from local E _t ). Located in the Western Sahel and influenced by the West African Monsoon (WAM) (Niang et al., 2020; Nouaceur & Murarescu, 2020), the basin is characterized by a distinct wet and dry season (Figure 5d). On average, only 12% (70 mm) of the annual P falls during the dry season (November–April). More than half of that derives from E _t (Figure 5e). Most of P falls in the wet season (on average 505 mm between May and October). Although most wet season P originates from the ocean, E _t still constitutes 36% of P, which implies that vegetation plays an important role to produce rainfall even when oceanic influx is dominant. Moreover, remote vegetation supplies most of the moisture for P during the onset and cessation of the wet season (light green line in Figure 5e). The alternating importance of oceanic (blue line) and remote E _t sources (light green line) in Figure 5e may suggest the existence of a land–atmosphere feedback in the post‐monsoon period to which vegetation‐sourced P contributes significantly (Breil et al., 2017; Notaro et al., 2019).

Over the years, we find an increase in mean annual P between 1981 and 2016 (Figure 6a). Similar trends were observed in earlier studies (Bodian et al., 2020; Nicholson et al., 2018; Nouaceur & Murarescu, 2020) which found that annual P has been recovering since the Sahelian drought of the 1970s, and linked the recent increase in P to rising SST over the Atlantic Ocean (Nouaceur & Murarescu, 2020). Correspondingly, we find an increasing importance of oceanic sources for P over Senegal for this period, with increasing relative and absolute contributions from the ocean (Figures 6f and 6g). This is in contrast to the decreasing contribution of oceanic evaporation to P in the Congo basin (Figure 4f), but in line with global studies (Findell et al., 2019; Gimeno, Vázquez, et al., 2020) that suggest an increasing importance of ocean evaporation for terrestrial P in light of climate change on a global average. CMIP5 models assessing 21st century precipitation characteristics of the WAM under RCP4.5 emission scenario also show increased rainfall during the monsoon season due to increased moisture flux convergence and local recycling (Monerie et al., 2016), although the predicted rainfall increase may be weakened by human‐induced land use and land cover change (Quesada et al., 2017). Furthermore, the importance of remote E _t seems to increase at the expense of local E _t (Figure 6c), although in terms of absolute contributions, both source contributions slightly increase over time (Figure 6b) and thus contribute to the wetting trend observed over the watershed (Figure 6a). At last, relative contributions of E _d sources appear to diminish (Figure 6e), although no clear trend appears in terms of absolute contribution (Figure 6d). Hence, the increase in oceanic evaporation sources seems the most significant driver of increasing rainfall patterns over the Senegal basin while being supported by an increasing contribution of vegetation‐sourced E _t on rainfall.

Figure 6.

(a) Annual P (mm) over the Senegal basin between 1981 and 2016; (b and c) Annual contribution of E _t to P in mm and fractional contribution to total rainfall, respectively. Full line represents local sources, dashed line represents remote sources; (d and e) Annual contribution of E _d in mm and fractional contribution to total rainfall, respectively. Full line represents local sources, dashed line represents remote sources. (f and g) Annual contribution of ocean E to P in mm and fractional contribution to total rainfall, respectively. Note the different scaling of the x‐axes corresponding to local (left y‐axis) and remote (right y‐axis) sources of E _t and E _d contributions.

3.3. Limitations

Despite the clear contribution of vegetation‐sourced P over the African continent, the level of analysis at the watershed‐scale applied here is likely to overlook regional variabilities. Considering the average size of the watersheds included in this study (∼1.2·10⁶ km²), spatial variability of rainfall and rainfall source regions within watersheds is likely to be high, at least for large watersheds covering various bioclimatic zones. Besides issues of scale, there are three major limitations:

First, although the results presented here indicate the importance of E _t in the continental water cycle over Africa, there are limited options to validate precipitation sources (Keune et al., 2022). The lack of ground‐based P data over Africa makes estimations and corrections of simulated rainfall patterns cumbersome (Washington et al., 2013). Yet, Sorí et al. (2017) applied a similar methodology—using FLEXPART moisture trajectories, driven by ERA‐Interim—and used ground‐based P and runoff measurements to construct the climatology for the Congo basin. They found a significant relationship between FLEXPART simulations and ground‐based data. However, the use of various precipitation products in studies over the Congo basin have led to contrasting conclusions regarding moisture source regions (Dyer et al., 2017) and precipitation trends (Washington et al., 2013).

Second, differentiation of E is based on daily E _t /E _d ratios from GLEAM, despite the lack of consensus on the best performing E product in Africa (Burnett et al., 2020; Miralles et al., 2016). Local comparisons with ground‐based flux‐tower meteorological data show that GLEAM has a generally good performance compared to other E models (see, e.g., McCabe et al., 2016; Michel et al., 2016). Over Africa, recent experiments on the performance of GLEAM have not reached converging conclusions (Dembélé et al., 2020; Khosa et al., 2019; Majozi et al., 2017). GLEAM estimates are at the high end of E _t /E _d estimates (Wei et al., 2017), which implies that E _t contributions to P, as presented in this study, are more likely to be over‐rather than underestimated.

Third, our presented moisture recycling metrics are strongly coupled to the choice of forcing and bias‐correction data (i.e., ERA‐Interim, OAFLUX, MSWEP, and GLEAM), as well as the water accounting method applied (i.e., Lagrangian moisture tracking using FLEXPART). For example, Keune et al. (2022) studied the uncertainty inherent in the evaluation of Lagrangian trajectories; according to their analysis, the present study may yield moisture recycling ratios that are on the upper end of model‐internal uncertainties. This finding is in line with other studies, using different datasets and models, and yielding lower moisture recycling estimates (see Table 1). However, as mentioned earlier, our results for the Congo align well with studies applying similar water accounting methods (Sorí et al., 2017), but are larger than those based on Eulerian frameworks (Dyer et al., 2017). Accordingly, the findings presented here are sensitive to the methodological choices, of which the strength and uncertainties could be addressed in further studies.

With respect to interpretations of the results, there are two major limitations. First, the overall P source regions show that vegetation‐sourced P is substantial over the African continent, which is corroborated by other studies that investigate the role of vegetation cover in moisture recycling (Aemisegger et al., 2014; Miralles et al., 2016; O’Connor et al., 2021; van der Ent et al., 2014; Wang‐Erlandsson et al., 2014; Zhao et al., 2019). Yet, these results cannot be directly used to infer the impact of land cover changes on P patterns. As also argued by Baudena et al. (2021), implicit assumptions of proportional change to P in response to changes in E _t imply a simplification. Changes in vegetation cover affect local coupling and atmospheric circulation via land–atmosphere feedbacks (Goessling & Reick, 2011)—which further influence moisture recycling dynamics. It remains unclear at which scale of land cover change this leads to certain effects on atmospheric circulation, although Lawrence and Vandecar (2015) suggest that a threshold for tropical deforestation exists at which rainfall is significantly reduced. Finally, the presented moisture recycling metrics are valid for individual watersheds, but are not comparable across watersheds, as the moisture recycling is per definition area‐dependent: the whole Earth has a moisture recycling ratio of 1, whereas an infinitely small area would have a moisture recycling ratio close to zero (Trenberth, 1999). As such, larger watersheds are expected to have higher moisture recycling ratios.

4. Conclusion

This study sets out to assess the spatial and temporal variability of rainfall sources over major African watersheds and identify the role of vegetation in contributing to these patterns. Our analysis shows that, on average, almost 50% of the rainfall over Africa is vegetation‐sourced, although the variation between watersheds is large: mostly coastal watersheds are predominantly dependent on oceanic evaporation, while many large inland watersheds—such as the Congo and Nile—show strong dependencies on vegetation‐sourced precipitation. Whereas some watershed show dependency on vegetation‐sourced precipitation year‐round (i.e., Congo and Niger), others show large seasonal discrepancies (i.e., Senegal and Zambezi). The two watersheds featured in this study highlight these contrasting patterns. In the Congo basin, local E _t is a significant contributor to year‐round P and also appears to sustain precipitation in surrounding watersheds. For the Senegal basin, seasonally varying dependencies on remote E _t suggests the existence of a vegetation‐feedback in the post‐monsoon period. Interestingly, we find contrasting P trends that can be attributed to trends in moisture contributions from the ocean for the two basins: over Senegal, increased rainfall corresponds to increasing contributions of oceanic evaporation, possibly linked to observed SSTs over the Atlantic as suggested in previous studies. However, over the Congo, a persistent drying trend corresponds to reduced contributions of oceanic evaporation, and increased relative contributions of local E _t to sustain rainfall. As such, this suggests that local vegetation‐sourced precipitation becomes increasingly important to sustain rainfall over the Congo. In light of rapid deforestation observed over Africa (FAO, 2020), it is crucial to understand how precipitation is affected by such interventions. The accompanying data set which this study presents can hence support further research on the contribution and importance of transpiration over different land cover types for rainfall trends and anomalies in the region.

Supporting information

Supporting Information S1

Te Wierik, S. A. , Keune, J. , Miralles, D. G. , Gupta, J. , Artzy‐Randrup, Y. A. , Gimeno, L. , et al. (2022). The contribution of transpiration to precipitation over African watersheds. Water Resources Research, 58, e2021WR031721. 10.1029/2021WR031721

Data Availability Statement

The data set supporting this manuscript (Te Wierik et al., 2022) is available via https://figshare.com/s/4a40bb2fe82b5db294c2, in NetCDF files at 1° resolution for individual watersheds (monthly precipitation source regions over 1981–2016). This includes the accompanying code to process and visualize the precipitation source regions. The data can be cited as: (Te Wierik et al., 2022, Rainfall sources over African Watersheds, Dataset. https://doi.org.10.6084/m9.figshare.17099888.v1). Other data used in this study: HydroSHEDS are available via https://data.apps.fao.org/catalog/iso/e54e2014-d23b-402b-8e73-c827628d17f4; ERA‐Interim data are available via https://apps.ecmwf.int/datasets/. GLEAM data are available via https://www.gleam.eu/; MSWEP data is available via http://www.gloh2o.org/. This publication is based upon the WHOI OAFlux datasets supported by the NOAA's Global Ocean Monitoring and Observing (GOMO) Program and NASA's Making Earth System Data Records for Use in Research Environments (MEaSUREs) Program. OAFlux data as used here, is available from https://oaflux.whoi.edu/data-access/. The framework for analysis of FLEXPART data (Keune et al., 2022) is available via https://github.com/h-cel/hamster.