Mapping groundwater recharge in Africa from ground observations and implications for water security

We developed a database of existing recharge studies for Africa. Recharge studies were compiled using published and grey literature identified from online databases, Google Scholar® and the Africa Groundwater Atlas (accessed May 2020) for the period 1970–2019, supplemented from the authors' personal databases and knowledge. For each study, meta-data including the time period of the recharge estimate, the geographic extent of the study, and an accurate geolocation checked using aerial photography from Google Maps®. Long-term average recharge (LTA recharge) values were recorded for each study area along with the employed methods noting the primary and secondary methods used. Each study was critically reviewed and given a confidence rating based on the number and type of methods applied, the quality of the application of the methods, and the number of sites included in the study (see supplementary data for more information (available online at stacks.iop.org/ERL/16/034012/mmedia)). A range of uncertainty was also included for each study: for studies with many measurements uncertainty was often the quoted standard deviation, interquartile range (IQR), or range; where studies noted uncertainty in input factors (such as specific yield or rainfall chloride) these were included in calculating the uncertainty. Estimates of groundwater recharge are inherently uncertain due to the assumptions made, therefore particular care was taken to exclude studies where the assumptions were invalid, and uncertainty was reduced when multiple independent methods were applied.

Different recharge-estimation methods included: CMB; water-table fluctuation (WTF) methods; environmental tracers (EnTs); WB; calibrated GMs; BF; and soil physics methods. Studies of groundwater recharge explicitly from river or lake infiltration, irrigation returns or urban leakage were excluded, leaving only studies representing distributed recharge from rainfall, which can include both diffuse and distributed focussed pathways. After the database was completed, studies for individual countries were assessed together to identify and remove duplicates; preference was given to syntheses and studies with higher confidence ratings. Studies with lower confidence were retained if no other study had been repeated at the location. Records in the final dataset had equal weighting in the analysis.

An analysis in a geographical information system (GIS) was then used to attribute various physical parameters known from previous studies to have an influence on groundwater recharge: long-term (1981–2010) climate data (Harris and Osborn 2020) (LTA rainfall, aridity index, number of wet days in a year, potential evapotranspiration (PET)); aquifer domain (MacDonald et al 2012), land cover (Arino et al 2007); the reference soil group as mapped in the Soil Atlas of Africa (Jones et al 2013), and normalised difference vegetation index (NDVI) (Didan et al 2015). Maps of these data are shown in the supplementary material. Each of these parameters were used in the linear mixed model (LMM) described below.

We required a statistical model which used a suitable subset of the measured covariates to predict LTA recharge at unsampled sites so that LTA Recharge could be mapped at continental scale. It was necessary to use an approach that was suitable for data with significant clustering (and therefore not an independent random sample) and that also selected a subset of predictor variables without over-fitting. To achieve this we used an LMM. Technical details are given in supplementary material and by Lark et al (2006). In an LMM a variable is modelled as a combination of fixed effects and random effects. The fixed effects define the expected value of the dependent variable given values of covariates (including categorical variables such as aquifer domain or soil group, and continuous ones such as LTA rainfall). The fixed effects therefore constitute a regression-type model.

Random effects in the LMM comprised a spatially dependent Gaussian random variable, and an independent error term (see supplementary material). The spatial term accounts for variability in LTA recharge observations around the fitted value, given that fixed effects are likely to be spatially dependent such that observations at nearby locations are expected to be more similar than those further apart. The independent error term accounts for factors such as error in the processes of estimating LTA recharge at a site, as well as any fine-scale sources of variation in the variable which are not resolved by the spatial distribution of the sample locations.

To fit the LMM, the parameters of the random effects were first estimated by residual maximum likelihood. The parameters of the fixed effects were then estimated by generalised least squares (Lark et al 2006). To select the final subset from predictors listed above, we considered predictors sequentially, starting with a simple 'null' model in which the only fixed effect was a constant mean value. A model with the first predictor was then fitted, and the strength of evidence that the predictor should be retained was assessed by a log-likelihood ratio (LLR) test following Welham and Thompson (1997). A continuous predictor would appear in the model as a regression-type predictor with an estimated coefficient. A categorical predictor (such as soil group) would appear in the model as a set of different intercept terms for the prediction. If a categorical predictor was accepted by the LLR test, then its interactions with continuous predictors already in the model would be tested next. Once all predictors had been considered, the evidence for each one included was reassessed against a threshold for the P-value determined to control the false discovery rate (FDR) over all tests conducted. The FDR is the probability that a predictor which has no relation to the target variable is incorrectly regarded as significant within a set of multiple hypothesis tests. We followed the approach described by Lark (2017) to maximise power to detect meaningful predictors while controlling FDR by ordering the initial sequence of predictors so that those thought most likely to be predictive were considered first.

The final model was then used to compute the empirical best linear unbiased prediction (E-BLUP) of LTA recharge at unsampled sites on a prediction grid. The E-BLUP combines the predicted value from the fixed effects with an interpolation of the random effect by a kriging-type predictor. The E-BLUP thus minimises the expected squared prediction error (prediction error variance), and computes this value as a measure of uncertainty. After an exploratory least-squares fit of a model with all predictors, and examination of the distribution of residuals, we transformed LTA recharge to natural logarithms to make the assumption of Gaussian errors plausible; continuous predictors were also transformed to logarithms because of their marked skewness. The prediction was back-transformed to the original scales of measurement by exponentiation, which gives a median-unbiased prediction, appropriate for skewed variables.

Average decadal groundwater recharge was subsequently calculated for each African country with uncertainty characterised by the prediction error variance of the best linear unbiased predictor. Groundwater recharge was then compared with published estimates of groundwater storage for each country (MacDonald et al 2012) to form a basis to discuss water security in terms of a combination of LTA recharge and storage (Foster and MacDonald 2014).

A total of 134 separate studies with robust quantitative LTA recharge data were identified for Africa. Figure 1 shows their location and the primary method employed in their analysis of recharge (MacDonald et al 2020). In arid, semi-arid and dry sub-humid areas, CMB was the most commonly applied method whereas EnT (such as tritium) and WTF methods were the most common methods in arid and humid areas respectively. However, 75% of the studies used two or more methods to estimate recharge, with CMB and EnT representing the most common combination. Studies are generally well distributed across Africa, though there is a relative dearth of studies in humid areas, such as the Congo basin, coastal West Africa, and southeastern Africa including Madagascar. Most arid and semi-arid areas are well represented with the exception of the Horn of Africa (figure 1).

Zoom In Zoom Out Reset image size

Measurable modern groundwater recharge was found in studies across much of Africa and rates decrease with aridity (figure 2). Although recharge is not detected in the majority of studies in hyper-arid areas, 40% of studies indicated some measurable recharge, and was generally associated with episodic rainfall and wadi infiltration (Edmunds et al 1992, Gonçalvès et al 2013). In arid and semi-arid climates, some component of modern groundwater recharge was detected in all studies, with median values of 60 mm (IQR 30–140 mm) and 200 mm (IQR 90–430 mm) respectively decade⁻¹, although recharge rarely occurred every year but was episodic (e.g. Taylor et al 2013b). In dry sub humid and humid areas groundwater recharge is significantly higher and recharge is recorded to occur during most years (figure 2): dry sub humid 920 mm (IQR 480–1600 mm) decade⁻¹; humid 1300 mm (IQR 760–2200 mm) decade⁻¹.

Zoom In Zoom Out Reset image size

Table 1 shows the results for successive addition of predictors in the fixed effects of the model. In row one of the table, a model with log_e LTA rainfall as the only predictor is tested against the null model with the only fixed effect a constant mean (µ). The LLR is a measure of the strength of evidence against the null model, with larger values reflecting stronger evidence. The P-value (≪0.0001) indicates the low probability that there is evidence to support the preference of the null model, therefore log_e LTA rainfall is retained. In the next row a model in which log_e PET is added as a second predictor is tested against the alternative model with LTA rainfall only. In this, and for all subsequent predictors considered, the evidence to retain the new predictor is weak. Because none of the categorical predictors were included, their interaction with LTA rainfall was not considered. Applying the FDA control, only LTA rainfall was retained as a predictor in the model. The final fixed effect model therefore takes the form of equation (1):

$$\begin{equation}{\text{lo}}{{\text{g}}_{\text{e}}}\left[ {{\text{LTA recharge}}} \right]\,{\text{ = }}\,{\beta _{\text{0}}}{\text{ + }}{\beta _{\text{1}}}{\text{lo}}{{\text{g}}_{\text{e}}}\left[ {{\text{LTA rainfall}}} \right]\end{equation} \tag{ 1 }$$

where β₀ is −5 and β₁ is 1.388 (table 2). Random effects parameter estimates are also shown for the overall null model where a constant mean is the only fixed effect and for the selected model. Spatially correlated random effects showspatial dependence up to distances approximately three times the distance parameter, φ (see supplementary material). Using a constant mean as the only fixed effect there is spatial dependence in the LTA recharge data at distances up to approximately 6000 km (reflecting the spatial dependence of long-term climatic trends). The spatial dependence of the random variation after including LTA rainfall as a predictor is restricted to a shorter range of 900 km. The term c₁ is the variance of this spatially correlated random effect. Note that c₁ is much reduced on inclusion of rainfall as a fixed effect. The term c₀ is the variance of the uncorrelated random error. The adjusted R² for the model, computed from the reduction in random variance on inclusion of the fixed effect, is 0.82. This is the proportion of the variance in log_e LTA recharge accounted for by including LTA rainfall in the model. The non-linearity in the relationship between LTA rainfall and LTA recharge represented by the model is consistent with ground-based studies highlighting the disproportionate contribution of heavy rainfall to groundwater recharge in Africa (e.g. Taylor et al 2013b, Cuthbert et al 2019a, Banks et al 2020).

The LMM was used to produce the map of LTA recharge (figure 3) along with observed data for comparison. The mapped value at any location is obtained as a combination of the predicted value obtained by applying the model in equation (1) to the known LTA rainfall values, and a kriging-type prediction from the residuals for this fit to minimise the expected prediction error variance. Figure 3(b) plots the relationship between LTA recharge and LTA rainfall along with the modelled relationship and data from Cuthbert et al (2019a) for comparison. Recharge estimates are more uncertain in the wetter areas of Congo basin, coastal West Africa and Madagascar as there are no recharge observations where LTA rainfall exceeds 2000 mm a⁻¹. Using this method the LTA decadal groundwater recharge for Africa is estimated to be 15 000 km³ (4900–45 000 km³). The range is the prediction error variance of E-BLUP.

Zoom In Zoom Out Reset image size

The countries with the lowest rates of distributed groundwater recharge flux are Western Sahara, Egypt, Libya, Algeria and Mauritania with less than 50 mm decade⁻¹, and the highest groundwater recharge fluxes (>1500 mm decade⁻¹) are found in Liberia, Equatorial Guinea, Sierra Leone, Gabon and the Republic of Congo. Figure 4 shows the LTA recharge volume for individual countries and table 3 reports the total recharge volume for each country along with the average decadal recharge depth.

Zoom In Zoom Out Reset image size

Combining estimates of LTA recharge with available groundwater storage provides a platform to discuss water security (Foster and MacDonald 2014) and the resilience of aquifer systems to drought (Calow et al 2010). This analysis can be undertaken at different scales, from individual aquifers to regions or countries. Figure 4(a) shows a map of combined groundwater recharge fluxes and groundwater storage for Africa and highlights high storage but low recharge in many of the North African sedimentary aquifers, and low storage but high recharge across many crystalline basement aquifers. Figures 4(b) and (c) compare groundwater storage and recharge for individual countries—with a combination of high groundwater storage and recharge giving increased water security. Countries in this situation are rare with only five countries, Guinea Bissau, the Republic of Congo, the Democratic Republic of Congo, Nigeria and Angola having both recharge and storage above the median for Africa. Having both low groundwater storage and recharge considerably reduces water security; again countries in this condition are rare with only five, eSwatini, Zambia, Lesotho, Zimbabwe and Eritrea having both recharge and storage below the African median.

The majority of African countries have either high storage, or high groundwater recharge. Where storage is high, but LTA recharge low, groundwater is resilient to short-term change in climate, but prone to long-term depletion. In Africa, six of the countries with least recharge have highest storage: Egypt, Algeria, Libya, Tunisia, Niger and Botswana and represent some of the highest groundwater abstractors in Africa (Wada et al 2012). In areas where groundwater storage is low but recharge is high, groundwater supplies are more vulnerable to drought but resilient to long-term irreversible depletion (figure 4). Five of the countries with highest recharge have the least groundwater storage: Liberia, Equatorial Guinea, Guinea, Cote D'Ivoire and Burundi. A similar picture of water security is given by considering the groundwater recharge and storage available per capita (table 3).

The LMM shows that at the scale of the African continent only LTA rainfall is related to LTA recharge. The inclusion of other factors, PET, NDVI, number of wet days, soil group, aquifer domain and land cover does not improve the model. This outcome contrasts with that of some other large-scale studies (e.g. Mohan et al 2017), which required several different climate and land-cover parameters to improve model fitting. The good fit of the model presented here using only LTA rainfall may be due to the extra care taken to geolocate studies and only include studies that passed stringent quality assurance tests to give data that estimate long-term recharge. However, the random effect in the LMM still show spatial dependence (up to 900 km), and local prediction of recharge can be improved by using the BLUP, with its kriging-type component, relative to a prediction from rainfall alone. This spatial dependence reflects the importance of local soil and land use conditions in mediating the relationship between recharge and rainfall as demonstrated by many studies (e.g. Sami and Hughes 1996, Scanlon et al 2005, Ibrahim et al 2014). However, in modelling LTA recharge across Africa using continental-scale datasets, these parameters do not have a consistent effect, and LTA rainfall dominates the relationship. The scale of these local effects is represented by the variance observed in the relationship between LTA rainfall and recharge (figure 3 and table 3). Therefore, the approach taken here has most value for estimating recharge at a regional scale.

There is little evidence from the data to support a threshold value of LTA rainfall below which recharge cannot occur. Measurable distributed modern groundwater recharge is found to occur where average annual rainfall is considerably less than 250 mm (figure 2) and even in some hyper-arid areas. This deduction is most likely explained by the increased probability of high rainfall events occurring when observing over several decades and the dominance of convective precipitation in arid areas. For example in the dune covered Stampriet Basin in Namibia where average rainfall is less than 250 mm a⁻¹, rainfall can be as high as 600 mm in some years—leading to significant diffuse recharge events (Stone and Edmunds 2012). Lack of vegetation in some arid and hyper-arid areas can also contribute to diffuse groundwater recharge, as heavy rainfall can infiltrate beneath the shallow zero flux plane (Dincer et al 1974, Gates et al 2008, Scanlon et al 2010). Focussed groundwater recharge is also of increased importance in arid areas (Scanlon et al 2006, Cuthbert et al 2019a) and can contribute to regional distributed groundwater recharge. In Niger for example, research has demonstrated highly localised redistribution of runoff by vegetation and geomorphology (Bromley et al 1997) and more widely distributed recharge through temporary ponds and ephemeral streams (Leduc et al 1997). The presence of karst feature can also increase focused recharge in arid regions (Schulz et al 2016). This current study examining long-term relationships complements recent findings examining patterns in annual recharge (Taylor et al 2013b, Kolusu et al 2019. Cuthbert et al 2019a), which demonstrate increasing annual variability in recharge with aridity and many years with no measurable observed recharge. These results strongly support measuring and reporting recharge values over decades rather than annually.

Most African countries with little groundwater storage have high annual rainfall and therefore regular recharge. Conversely, many African countries with low rainfall, usually considered as water insecure, have considerable groundwater storage which was mostly recharged millennia ago (figure 4). Generally, assessments of water stress and water security ignore groundwater storage (Taylor 2009), but concentrate on annual ratios between renewable water and water use by people and the environment (Vörösmarty et al 2010). Even where groundwater is considered, early thinking was dominated by the concept of safe yield and ratios between abstraction and recharge (Zhou 2009). However over the last 2 decades this conceptualisation has evolved beyond simple ratios to broader issues of sustainability (Morris et al 2003, Alley and Leake 2004); and include both the response time of groundwater systems to change (Cuthbert et al 2019b) and dynamic groundwater storage (Damkjaer and Taylor 2017, Bierkens and Wada 2019). Sustainability is increasingly being considered as context specific (Gleeson et al 2020) with an acknowledgement that countries have different priorities depending on other pressures and aims, such as the Sustainable Development Goals. With increasing calls to draw from groundwater storage in order to stimulate economic growth and improve food security in Africa (Wijnen et al 2018, Cobbing 2020) a more nuanced approach to water security is necessary.

Several countries, particularly, (but not exclusively) in North Africa (figure 4(c)) have considerable water security when groundwater storage is taken into consideration. This storage provides a significant buffer before abstraction will impact the regional groundwater system. Since there is little dependency on modern recharge, the supply is secure against current climate variability. Examples from elsewhere, including the High Plains and Californian Central Valley aquifers in the US (Scanlon et al 2012), as well as the Indo-Gangetic Basin in northern India and Pakistan (MacDonald et al 2016) demonstrate that such abstraction can be continued in large aquifers but not without degradation. Similar degradation occurs in smaller heavily exploited aquifers North Africa, such as the Souss in Morocco (Bouchaou et al 2008). In Libya, where groundwater recharge is negligible, groundwater storage was developed as part of the Great Man Made River Project without convincing evidence of its sustainability (Edmunds and Wright 1979, Ahmad 1983). Unsustainable groundwater abstraction increased current water supply considerably but at the expense of water security for future generations. Eventually limits can be reached as abstraction increases and becomes more widespread. As groundwater-levels fall, and groundwater discharge reduces in response to abstraction, ecological consequences can occur and include the destruction of wetlands, reduction of baseflow to rivers, land subsidence and increases in salinity. Equally important is the impact on other groundwater users as shallow wells and boreholes dry up, leaving only those with access to the deepest and most expensive boreholes able to access groundwater.

Countries with low groundwater storage but high groundwater recharge are common in Africa, mainly because of the wide prevalence of crystalline-rock and volcanic aquifers. When abstraction is low (e.g. supporting widespread community hands pumps) the water supply is secure, even through short-term droughts (MacDonald et al 2019). Where aquifers are more permeable however, the increase in abstraction (enabled by the geology) for irrigation or urban supply, can quickly deplete storage in dry years—as shown in the crystalline basement aquifers in Peninsular India (Collins et al 2020).

Both these examples above highlight the importance of ongoing dedicated groundwater monitoring to act as an early warning of rapid changes in storage or water quality in response to pumping. Although the use of GRACE has proved useful at scales of >100 000 km², it cannot substitute for in situ monitoring to capture changes in groundwater levels (Shamsudduha and Taylor 2020), water quality (Lapworth et al 2020) and abstraction (Oiro et al 2020) to identify threats to water security for individual aquifers, regions or countries.

The three most common methods used to estimate LTA recharge in Africa are CMB, WTF methods and the use of EnTs. The majority of studies used several techniques to assess recharge, and the most reliable studies (those in the Sahel and the Kalahari) have developed understanding and data over decades, with different methods and new techniques used to challenge and refine previous results. There is no evidence of bias in the dataset from the method used (see supplementary material). This observation likely arises, however, from the quality assurance procedure, which eliminated studies where techniques had been applied incorrectly and gave a lower priority to studies where only one technique was applied.

Where correctly applied (e.g. accounting for surface runoff, and constraining rainfall and atmospheric inputs) CMB offers useful insight for integrating recharge over the long-term. Studies where unsaturated zone profiles of chloride concentrations are combined with measurements at the water -table are particularly helpful (e.g. de Vries et al 2000, Edmunds et al 2002), as are the additional use of EnTs in more arid areas to indicate if recharge has occurred within recent decades (Lapworth et al 2013). The WTF method is the only method which directly measures groundwater storage responses and is well suited for examining temporal responses to recharge. However to estimate LTA recharge in more arid areas a sufficiently long record is required to capture important episodic recharge events (Taylor et al 2013b, Cuthbert et al 2019a) and sufficient local knowledge is needed of the aquifer properties and the groundwater discharge regime (Cuthbert 2010). In humid areas however, where annual recharge is less variable (Cuthbert et al 2019a), shorter hydrographs can also provide valuable data particularly when in combination with other methods. In arid and hyper-arid areas, combinations of EnTs are used to estimate the residence times of groundwater, and in particular identify any component of modern recharge.

There are large parts of Africa including 19 countries lacking a reliable recharge study. In humid areas of Central Africa, West Africa and Madagascar, there is a particular dearth of studies, probably due to high rainfall and limited reliance on groundwater. The limited understanding of recharge process and volumes in these areas has consequences when modelling continental and global processes and constraining terrestrial WBs. In more arid areas, the Horn of Africa is lacking in studies—an area where groundwater is fundamental for buffering drought (MacAllister et al 2020).