Sustained Water Storage in Horn of Africa Drylands Dominated by Seasonal Rainfall Extremes

Rural communities in the Horn of Africa Drylands (HADs) are increasingly vulnerable to multi‐season droughts due to the strong dependence of livelihoods on seasonal rainfall. We analyzed multiple observational rainfall data sets for recent decadal trends in mean and extreme seasonal rainfall, as well as satellite‐derived terrestrial water storage and soil moisture trends arising from two key rainfall seasons across various subregions of HAD. We show that, despite decreases in total March‐April‐May rainfall, total water storage in the HAD has increased. This trend correlates strongly with seasonal totals and especially with extreme rainfall in the two dominant HAD rainy seasons between 2003 and 2016. We further show that high‐intensity October‐November‐December rainfall associated with positive Indian Ocean Dipole events lead to the largest seasonal increases in water storage that persist over multiple years. These findings suggest that developing groundwater resources in HAD could offset or mitigate the impacts of increasingly common droughts.

across the region as we write. Given the precarious relationship between climate, the water cycle, and human lives and livelihoods in the HAD, there is a need to assess the sustainability of various components of the water balance over recent decades.
Over the last two decades in East Africa drylands, there have only been three MAM rainy seasons above the centennial average, while droughts occurred in nine out of the 20 years (Funk, Shukla, et al., 2019;Funk et al., 2018). This suggests that the "new normal" in HAD is characterized by lower-than-average MAM rainfall and rising drought frequency during a key growing season for HAD. Meanwhile, increasing rainfall in the OND season has been documented, with high interannual variability linked to the phase of the Indian Ocean Dipole (IOD; Liebmann et al., 2014;Nicholson, 2017), as well as to higher frequency of major flood events such as in 1997 and those that affected Somalia and Kenya in 2019 (FEWSNET Special Report, 2020). Despite documentation of these recent climatic changes, the impacts of MAM and OND rainfall on subsurface water storage across the HAD are currently unknown due to the lack of in situ observational networks for SM and groundwater. Remote sensing data have been used to estimate changes in parts of East Africa on local scales, over limited timeframes, or as part of global compilations (Bonsor et al., 2018;Kolusu et al., 2019;Shamsudduha & Taylor, 2020), but the broader picture of trends in spatial variability in water storage in response to seasonal rainfall across HAD is not known.
Drought propagation is poorly understood in drylands (Calow et al., 2010;MacAllister et al., 2020;Quichimbo et al., 2020). In dryland regions, hydrological partitioning of rainfall into infiltration, evapotranspiration, runoff, and groundwater recharge is highly sensitive to the spatial and temporal distribution of rainfall and most particularly, to the intensity and duration of individual rainfall events (Apurv et al., 2017;Cuthbert et al., 2019;Kipkemoi et al., 2021;Mileham et al., 2009;Taylor et al., 2013), rather than simply to total seasonal or annual rainfall. This is because the replenishment of groundwater ("recharge") in drylands occurs mainly via focused recharge (transmission losses of flowing water into streambeds due to the predominance of surface runoff, Quichimbo et al., 2020) as well as by diffuse recharge (deep infiltration of water through soils). Sustained subsurface water storage in regions such as HAD therefore, depends on rainfall being of sufficient intensity to generate deep local infiltration and also surface runoff over the landscape that is routed to and recharged through porous ephemeral channels.
Trends in seasonal rainfall totals tend to be the focus of most HAD climate studies due to their importance for SM drought conditions during the primary MAM crop planting season (e.g., Agutu et al., 2017). Past work has revealed declining MAM and increasing OND rainfall (Harrison et al., 2019;Liebmann et al., 2014;Tierney et al., 2015), which might be expected to affect water storage. Here we explore the impacts of seasonal rainfall changes on water storage across HAD based on corroboration from multiple gridded data sets. Thus, we aim to address the causes and effects of water storage changes where there is a notable lack of in situ observational networks, and thereby make progress in understanding a consequential problem for the region. We (a) compare trends in seasonal rainfall within widely used rainfall data sets; (b) analyze decadal trends in Gravity Recovery and Climate Experiment (GRACE)-derived total water storage (TWS) and in shallow SM via the European Space Agency's Climate Change Initiative (ESA-CCI); (c) determine the dominant seasonal contributions to the TWS and SM signals via seasonal differencing; and (d) explore correlations between rainfall trends (in totals and extremes) and water storage trends across HAD. Our analysis is designed to generate new understanding of how HAD climate variability and trends have affected subsurface water storage in recent decades, and to address whether groundwater may potentially provide a robust resource for mitigating regional food and water insecurity.

Methods
Rain gauges with continuous records spanning multiple decades are rare and sparse in HAD. Assessments of rainfall anomalies and trends therefore often rely on gridded data based on: spatially interpolated rain gauges; a blend of point gauge data with remote sensing; reanalysis climate model data; or a combination of all these methods to assess corroboration between them: We selected six commonly used rainfall data sets representing this range of methods: GPCC, CRU, CHIRPS, ARC2, ERA5, and MSWEP (Text S1 and Table S1 in Supporting Information S1). We quantified multi-decadal trends in monthly and seasonal (MAM and OND) rainfall over the 1983-2016 period that is common to all data sets (Figure 1a), averaged over a 1 × 10 6 km 2 area of HAD that exhibits bimodal rainfall in these seasons (black footprint in Figure 2a; Supporting Information S1; Nicholson, 2017). We took this multi-data set approach to identify consistent seasonal rainfall trends in HAD, because it has been shown that climate data sources have different strengths and weaknesses based on underlying methods and spatial and temporal resolutions. We then selected two widely used data sets with sub-monthly, CHIRPS , and sub-daily temporal resolution, MSWEP (Beck et al., 2019), to explore regional patterns of trends in seasonal and annual rainfall across HAD. The CHIRPS data set benefits from incorporation of a larger number of gauge observations in Ethiopia and Somalia, while the MSWEP gridded product has the advantages of high temporal resolution (3 hr) and incorporation of reanalysis climate data (Supporting Information S1).
In order to further understand the impact of rainfall on HAD water storage, we calculated trends in rainfall extremes (defined here as 3-hourly rainfall magnitude >95th percentile within MSWEP), and we compared these rainfall trends with secular trends in TWS from GRACE and in SM from the passive SM product of the ESA-CCI (integrated over an assumed depth of 10 cm, ESA-CCI 2021, https://www.esa-soilmoisture-cci.org/ sites/default/files/documents/public/CCI%20SM%20v06.1%20documentation/ESA_CCI_SM_D4.1_v2_PVIR_ v6.1_issue_1.0.pdf; Section 2) for the years of overlap between these data sets (2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016). October-November-December (OND; short rains) from six datasets. Each distribution represents the monthly rainfall trend across all pixels in the area framed black in panel (b). All trends are calculated over the period 1983-2016, depicted are the median and the 7th, 13th, 25th, 75th, 87th, and 93rd percentiles. Diamonds represent outliers. (b) Linearly approximated time series trends in annual, MAM and OND rainfall in HAD between 1981 and 2016. Grid cells with non-significant trends are framed in gray. (c) Time series of extreme seasonal rainfall total from MSWEP, 1979-2019 (falling in 3-hourly periods above the 95th percentile) in four equidistant locations across HAD, along with the approximated linear trend based on the Theil-Sen estimator (Supporting Information S1). The data for each location is averaged over an area of 1.5° × 1.5°.
For the correlation between TWS and SM with recharge with rainfall anomalies, we accounted for the initial state of TWS before each rainy season by computing TWS changes associated with MAM and OND seasons each year by quantifying the changes from the month before each season to the month after it, which we denote as ΔTWS (Supporting Information S1). This method removes autocorrelation in the TWS signal and the overall TWS trend and ensures that we are computing TWS contributions arising only from seasonal rainfall by bracketing rainy seasons with typically dry periods. We assume that these dry-to-dry season differences will fully incorporate into the ΔTWS signal the infiltrated rainfall and groundwater recharge occurring during a rainy season ( Figure S5 and Text S1 in Supporting Information S1). Accordingly, ΔTWS MAM was computed as TWS Jul -TWS Feb , and ΔTWS OND was computed as TWS Feb -TWS Sep , and we used the accumulation of these increments to calculate trends in ΔTWS associated with each rainy season over the GRACE period. We calculated seasonal SM changes by the same differencing method, denoted as ΔSM. The choice of period of differencing enabled us to capture the change in dry-season values of TWS before and after MAM and OND rainy seasons. ΔTWS OND for 2013 and ΔTWS MAM for 2014 were not available because of data gaps in GRACE, so we linearly interpolated TWS for the missing months based on the TWS values of the previous and following months. We then estimated the extent to which ΔTWS trends are associated with shallow SM versus deeper water storage (e.g., deep SM and groundwater).
Seasonal values and trends in total rainfall, rainfall extremes, ΔTWS and ΔSM were computed for all 0.5° pixels within the bimodal rainfall region (Figure 1c). To generalize these trends across the region, we spatially averaged over four roughly equidistant locations of 3 × 3 pixels (22,500 km 2 ) within HAD, located in Kenya, Ethiopia, Somalia and Somaliland (Figure 1c), based on the assumption that at an averaged scale of 1° and above, signal-to-noise ratios provide reasonable results for extreme rainfall (e.g., Roca, 2019). Our spatially averaged boxes of 1.5° × 1.5° thus allowed us to examine the impact of localized extreme precipitation, while not over-emphasizing local (single pixel) precipitation estimates. We also analyzed rainfall and water storage trends for all the pixels in this region.  (Amrhein et al., 2019), the average value of this trend across the HAD region is +4 mm/yr, which is consistent with that documented in Scanlon et al., 2022. (b) Bootstrapped overall trend and seasonal contributions to the TWS trend for the four locations indicated in panel (a) and for HAD as a whole. This subplot shows the overall linear trend in monthly TWS (All), and the contributions from long rains (March-April-May, MAM) and short rains (October-November-December, OND), based on accumulated ΔTWS for each season (see "Creating ΔTWS and ΔSM" in Supporting Information S1); outer and inner box boundaries represent the 5th-95th and 25th-75th percentiles of the bootstrapped trend contribution estimates and the central dot indicates the median value.

Results
Our results show broad agreement between rainfall data sets, portraying negative MAM and positive OND rainfall trends over the period of analysis (Figure 1a and Figure S1 in Supporting Information S1). The consistency between data sets based on different methodologies, including a reanalysis product that does not include rain gauge data in HAD (ERA5), suggests that these trends are not an artefact of progressively decreasing numbers of rain gauges for the region, as has been suggested (Lorenz & Kunstmann, 2012;Maidment et al., 2015). The negative MAM trend is mostly due to total rainfall reductions in April and May, consistent with the documented trend of an earlier cessation of the long rains (Wainwright et al., 2019), while the positive OND trend is associated with higher rainfall in October and November (Figure 1a).
Our four selected locations across HAD exhibit both positive and negative trends in extreme MAM rainfall, but only positive trends in rainfall extremes for the OND season (Figure 1c and Figures S2 and S3 in Supporting Information S1). Averaged over HAD, seasonal rainfall totals decreased in MSWEP by approximately −0.9 mm/ yr (range across rainfall data sets: −1.6 to −0.5 mm/yr) since 1981 while OND rainfall increased by 1.14 mm/yr (0.0 to 1.8 mm/yr) over the same period in MSWEP. MSWEP has the sufficient temporal resolution to identify high intensity rainfall events, and shows that these changes are largely due to changes in extreme rainfall, where rainfall above the 95th intensity percentile declined by −0.75 mm/yr in MAM, while it increased by 0.94 mm/ yr during OND. Additionally, in the last two decades HAD received on average as much rainfall during OND (123 ± 109 mm) as during MAM (124 ± 49 mm), despite the fact that OND has historically lower seasonal rainfall totals. Furthermore, the trends in extreme rainfall over this period are larger than trends in non-extreme seasonal rainfall, highlighting that recently a larger proportion of seasonal rainfall is occuring during extreme events than in the 1980s for both rainy seasons (Figures S3c and S3d in Supporting Information S1).
Over the GRACE period (2003-2016), we found that TWS exhibits a positive trend of ∼4 mm/yr (Figure 2a, Figure S6 and Table S2 in Supporting Information S1), suggesting that despite the increasing frequency of droughts and failing MAM rains (Figure 1; Funk, 2012), subsurface water storage has not declined, and is in fact rising across HAD. This rise in TWS based on contributions from both rainy seasons is most evident in inland areas of Kenya and in Ethiopia within the study area and less apparent over the southern coastal area of Somalia (Figure 2b). The short record and high interannual variability precludes statistical significance of these rising TWS trends (Supporting Information S1; Figure S7 in Supporting Information S1, Amrhein et al., 2019). Nevertheless, a bootstrapping analysis ruled out any significant negative trend in TWS (Supporting Information S1), and we suspect statistically significant positive TWS trends will develop in the future. For northern Kenya and southern Somalia (locations A and B), we find that the annual TWS storage trends are comprised of positive ΔTWS from both seasons, while for Somaliland (location D), negative ΔTWS MAM is offset by strongly positive ΔTWS OND , producing an overall positive annual trend in TWS (Figure 2b). In contrast, positive ΔTWS MAM compensates for small negative ΔTWS OND in eastern Ethiopia (location C), leading to an overall positive trend in annual TWS (Figure 2b).
By investigating the pixel-by-pixel correlations between seasonal GRACE ΔTWS trends ( Figure 2b) and seasonal rainfall trends derived from MSWEP across the HAD region (Figure 3; see Figure S7 in Supporting Information S1 for the same analysis with CHIRPS), we find that ΔTWS is positively correlated (r >0.5) with rainfall in both rainy seasons across most of HAD. However, there is high spatial heterogeneity and correlations are generally stronger in the southwest of HAD than in the northeast (Figure 3c), where the signal-to-noise ratio in rainfall estimates is high due to the higher aridity and lower seasonal totals. Across HAD, the correlation between ΔTWS MAM and MAM rainfall anomalies is slightly higher on average than between ΔTWS OND and OND rainfall, but the OND correlations are more spatially coherent. Annual values of ΔTWS OND and OND rainfall reveal the coincidence of positive IOD events (2006, 2011, and 2015), with high rainfall across HAD and large positive ∆TWS. In these positive phase IOD years, most of the four locations in Figure 2a received above average amounts of extreme rainfall (Figure 1c), suggesting the importance of these rainfall extremes in supporting sustained water storage.
We further ascertained how much of the ∆TWS variability can be explained by extreme rainfall by comparing the partial correlations between ∆TWS and total rainfall separately from extreme rainfall (>95th percentile). Even though on average the extreme part of the distribution contributes to around half of the seasonal total rainfall, we found that these extremes explain a disproportionate amount of the variance in ∆TWS. Specifically, rainfall in the >95th percentile contributes to >10 fold of the variance than the ≤95th percentile across most of HAD during OND, and across central Kenya and the Kenya-Somalia borderland during MAM ( Figure S9 in Supporting Information S1).
In general, GRACE TWS represents all surface water, SM in shallow layers, deep SM, and groundwater (e.g., Wahr et al., 2004). Since the HAD region (east of the African Great Lakes) largely lacks surface water because of its semi-arid/arid climate ( Figure S5 in Supporting Information S1), that leaves SM, deep SM, and groundwater in the TWS signal. We find that ∆SM values across HAD are at least an order of magnitude smaller than ∆TWS values for both rainy seasons (Figure S11a in Supporting Information S1), and that the correlation between seasonal rainfall and ∆SM is spatially restricted for both seasons (Figure S11b in Supporting Information S1). Finally, we find zero or negative trends in SM across the HAD region ( Figure S11c in Supporting Information S1). Taken together, these results lead us to conclude that SM comprises a very small part (if any) of the TWS signal across HAD, strongly suggesting that annual and seasonal TWS trends are dominated by water storage changes in deeper soil layers (below 10 cm) and in groundwater aquifers. Figure 4 summarizes all our trend results for the (c) Pixel-by-pixel correlations between seasonal total rainfall and ∆TWS for MAM and OND (stippling applied where correlations fall below the 95% significance level). (d) Same as (c) but with seasonal rainfall above the 95th percentile rather than total rainfall. In panels (a and b), the MAM years with the strongest zonal Pacific gradient are outlined with a red border and OND years with the strongest positive IOD years are outlined with a blue border.
various parts of HAD and averaged for the HAD as a whole, in terms of total seasonal rainfall, extreme seasonal rainfall, TWS, SM, and the contributions to ΔTWS from the ratio of extreme (r >95) to non-extreme (r <95) rainfall, where arrows represent the direction of trends and thickness indicates trend strength.

Discussion
In this paper, we investigated trends in seasonal rainfall within the HAD and their potential impact on regional water storage. The six rainfall data sets analyzed here corroborate previously documented seasonal rainfall trends in HAD for recent decades, showing that MAM rains have declined (Funk et al., , 2008Liebmann et al., 2014;Lyon & DeWitt, 2012;Williams & Funk, 2011). This multi-decadal decline is most strongly expressed in the latter part of the MAM rainy season, suggesting a possible shift in rainfall timing and/or a weakening of the climate drivers of rainfall in this season Nicholson, 2017;Wainwright et al., 2019). Our analysis also showed a stronger and positive trend in OND rainfall, wherein seasonal totals have increased by 0.5-1.0 mm/yr in both October and November (Figure 1a), punctuated by positive IOD events that contributed to recent large regional floods and are projected to occur more frequently in the future (FEWSNET Special Report, 2020;Funk et al., 2018). Furthermore, we found that TWS has increased across HAD, in agreement with a recent analysis of trends in the Ogaden-Juba aquifer (Scanlon et al., 2022).
Over the past 20 years, there have been ocean-atmosphere teleconnections affecting the HAD rainfall Liebmann et al., 2014). These include a zonal SST gradient in the Pacific, associated with low rainfall (drought) in MAM (Funk, Pedreros, et al., 2019;Funk et al., 2018) and a clear link between the IOD and high OND rainfall (MacLeod et al., 2020;Nicholson, 2017). Our research bears out these phenomena and their potential impacts on water resources across the HAD. We identify a link between positive IOD years (2006, 2011, and 2015), high OND rainfall, and large positive ∆TWS (Figures 3a and 3b), as well as particular years of low MAM rainfall associated with the zonal SST gradient in the Pacific , 2011Funk et al., 2018), leading to small and/or negative ∆TWS (Figures 3a and 3b).   Figure 2a) and averaged over the entire HAD. Trends are shown where the linear slope on the trend exceeds 0.1 mm/yr. Also shown is the ratio of variance in ∆TWS explained by extreme (r > 95) versus non-extreme rainfall (r < 95) from Figure S9 in Supporting Information S1, where red arrows indicate ratios <1, blue arrows show ratios >1, and thick blue arrows indicate ratios >2 and thick red arrows ratios <0.5. Trend values and correlation coefficients of the variables with time are listed in Table S2 in Supporting Information S1. National boundaries are shown in gray, and the political boundary of Somaliland as a gray dashed line. likely to be driving the correlation between seasonal rainfall extremes and ΔTWS. In other words, any recharge of deep SM and groundwater from one dry season to the next is likely controlled by the most intense rainstorms of the intermediate rainy period, forcing deep local infiltration across the landscape (diffuse recharge) and producing runoff that becomes streamflow in channels. This streamflow subsequently infiltrates through porous channel beds (focused recharge) that are typical of dryland regions, thus supporting regional groundwater aquifers (Cuthbert et al., 2019;Jaeger et al., 2017;. Recent research highlighted the renewability of groundwater storage across much of Africa (MacAllister et al., 2020;McDonald, 2021), but it lacked in situ groundwater data for HAD, one of the most food-and water-insecure and data-sparse regions in Africa. Our findings support the mechanism of deep-water recharge by extreme rainfall across HAD, contributing to a water store that persists and accumulates over decades, despite frequent droughts.
An important question is how seasonal rainfall in HAD will evolve in the future and what impacts it will have on water storage. Although CMIP3 and CMIP5 climate model simulations consistently project increased rainfall in both rainy seasons (e.g., Ongoma et al., 2018), the "East Africa Climate Paradox" highlights major concerns within the climate community about the veracity of these increasing projections, particularly for MAM in the face of observed negative rainfall trends from gridded data (Nicholson, 2017;Wainwright et al., 2019). These model projections for a wetter MAM may be in part based on theoretical increases in rainfall intensity for HAD, due to atmospheric warming via the Clausius-Clapeyron relation, leading to rainfall delivery in fewer, but more intense events (Chadwick et al., 2015;Trenberth, 2011). However, the magnitude of the MAM rainfall changes and the responsible dynamics within the models are unclear (Rowell, 2019;Shongwe et al., 2011). For OND, strong ENSO (Cai, Borlace, et al., 2014) and IOD events (Cai, Santoso, et al., 2014;Tierney et al., 2015) are projected to increase under anthropogenic climate change promoting more intense rainfall. Although higher intensity rainfall is associated with increased flood risk, the silver lining is that the increased channel runoff (floods) leads to higher rates of focused groundwater recharge with potential benefits for water security across the HAD region (Figures 3 and 4).

Conclusions
Our analysis of rainfall and water storage data sets showed that while total seasonal rainfall has declined in HAD over recent decades, especially during the agriculturally important MAM rainy season, TWS increased over the period 2003-2016. Furthermore, we showed that seasonal TWS dynamics are correlated with rainfall anomalies, and that positive trends in rainfall above the 95th intensity percentile is the major signal in this correlation. The ongoing trend of increasing water storage across HAD is potentially significant for drought adaptation and sustainability of water resources in this region. Economic and food security of HAD households are tightly linked to agricultural and pastoral livelihoods that depend on water availability under a changing climate. As of July 2022, drought-related food insecurity in Kenya and Somalia has pushed 18.6-21.1 million people into extreme (pre-famine) food insecurity (FSNWG Drought Special Report July, 2022). Hence, the link between increasing groundwater and rainfall extremes, described here for the HAD, may provide a valuable path towards regional climate adaptation and disaster risk reduction (United Nations, 2011). For example, a recent case study from Ethiopian drylands showed that groundwater may be a feasible water source for irrigation (Gowing et al., 2020). Given the dependence of rainfed agriculture on seasonal rains, which have repeatedly failed in recent decades, higher availability of groundwater for irrigated agriculture could potentially buffer the negative impacts of drought. The potential usefulness of groundwater depends on sufficient awareness of this water storage, its quality, accessibility as well as investment into relevant infrastructure to exploit it. In 2019/2020 this region experienced extreme precipitation, which was then followed by poor OND/MAM 2020/21, OND/MAM 2021/2022 rainy seasons, and as we write, a multi-year La Nina threatens to bring low rains again in OND 2022. It is possible that the stored water from the 2019/2020 rains could provide water resources to help buffer these more recent deficits.
Our analysis also suggests that it may be straightforward to identify parts of HAD where groundwater will become more abundant in the future, especially since the larger inter-annual variability implies that multi-year droughts are less likely for both rainy seasons in a row . Such an analysis, along with investment in an observational network of high-resolution rain gauges, SM sensors, and groundwater monitoring wells, could be used to identify cost-effective investment in local sand dams and even groundwater pumping equipment for broad regional benefit (MacDonald et al., 2009). Our findings suggest that groundwater in HAD is sustainable under the prevailing climate, and that rural subsistence communities of the HAD may have an opportunity to become more resilient to climatic shocks in the future based on increasing and sustainable groundwater reserves.

Conflict of Interest
The authors declare no conflicts of interest relevant to this study.