‘Eastern African Paradox’ rainfall decline due to shorter not less intense Long Rains

An observed decline in the Eastern African Long Rains from the 1980s to late 2000s appears contrary to the projected increase under future climate change. This “Eastern African climate paradox” confounds use of climate projections for adaptation planning across Eastern Africa. Here we show the decline corresponds to a later onset and earlier cessation of the long rains, with a similar seasonal maximum in area-averaged daily rainfall. Previous studies have explored the role of remote teleconnections, but those mechanisms do not sufficiently explain the decline or the newly identified change in seasonality. Using a large ensemble of observations, reanalyses and atmospheric simulations, we propose a regional mechanism that explains both the observed decline and the recent partial recovery. A decrease in surface pressure over Arabia and warmer north Arabian Sea is associated with enhanced southerlies and an earlier cessation of the long rains. This is supported by a similar signal in surface pressure in many atmosphere-only models giving lower May rainfall and an earlier cessation. Anomalously warm seas south of Eastern Africa delay the northward movement of the tropical rain-band, giving a later onset. These results are key in understanding the paradox. It is now a priority to establish the balance of mechanisms that have led to these trends, which are partially captured in atmosphere-only simulations. Springtime drying trends in East Africa can be explained by shifts in rainfall seasonality and not reductions in daily rainfall. The East African “Long Rains” — which fall during March, April and May — deliver the majority of precipitation to the region. Since the mid-1980s, observations reveal a reduction in rainfall which some climate models fail to capture. Using a suite of observations, reanalyses, and atmospheric simulations, Caroline Wainwright from the University of Reading, UK, and colleagues investigate the mechanisms behind this change. The reduction in Long Rain precipitation is characterised by a later onset and earlier cessation — and thereby a shortening — of the rainy season, attributed to a more rapid movement of the rain band associated with warmer SSTs off East Africa. Understanding these mechanisms will benefit climate projections and adaptation measures in an already-vulnerable region.


INTRODUCTION
The March, April and May (MAM) 'Long Rains' of Eastern Africa have historically been the major rainfall and primary agricultural season for much of the region 1,2 .Since around 1985 the rains have declined, with major consequences for livelihoods [1][2][3][4][5][6] .In contrast, climate model projections show increased long-rains rainfall: this has been termed the 'Eastern African climate change paradox' that has been used to question the reliability of the projections 5,7 and so arguably restricts their utility for informing suitable adaptation measures 6 .
Shortage of in situ observations makes assessment of regional rainfall trends across Africa challenging, yet the recent drying of the Eastern African long rains appears robust across datasets 4,6,8 , with a drying in March, April, and May 6 .Coupled Model Intercomparison Project Phase 5 (CMIP5) simulations fail to consistently capture the MAM Eastern African drying with around half simulating a drying, and half a wetting 6 , while most atmosphere-only (AMIP) simulations, driven by observed sea surface temperature (SST) and realistic radiative forcings, correctly capture the drying 3,6 .The observed drying is thus partly explained by SSTs 2,3,9 .Comparison of the observed long rains trend with a range of trends from control runs of CMIP5 coupled models, in which anthropogenic forcing (and external natural forcing) are absent, indicates that the observed rainfall trend is unlikely to be entirely consistent with natural variability 6 .Mechanisms have been proposed linking the decline in the long rains with decadal variability in the Pacific Ocean [1][2][3][4][5] .Specifically, changes in the zonal SST gradient in the Pacific 10 , warming of the Indo-Western Pacific SSTs (with anomalous warm water transported from the tropical western Pacific to the Indian Ocean by Indonesian throughflow 11 ) and enhanced convection over the Western Equatorial Pacific, are associated with an anomalous Walker circulation over the Indian Ocean.Strengthening of the upper level easterlies and the descending branch lead to increased subsidence over Eastern Africa and reduced precipitation 2,5,7 .However, there is limited evidence for the Indian Ocean Walker Cell extending over the Horn of Africa 1 , and an equatorial zonal circulation cell is not present in boreal spring 12 , although modulation of subsidence over Eastern Africa by upper level easterlies over the Indian Ocean has also been proposed 1 .
Projections of changes in the onset and cessation of wet seasons across Africa under future climate change, produced using CMIP5 models, show the onset and cessation of the Eastern African October-December "short rains" getting later, while the cessation of the long rains is projected to get earlier 13 .Changes in the timing of onset and cessation across interior Africa (0-35°E) have been linked to changes in the progression of the tropical rain-band and the Saharan Heat Low (SHL) 13 .A stronger SHL (amplified by a water vapour-greenhouse gas warming feedback 14,15 ), leads to a northward shift in the tropical rain-band during the boreal summer and a delayed retreat of the rain-band southwards 13 .This delayed retreat may be associated with the later onset and cessation of the short rains.
Here, specific characteristics of the long rains decline are investigated, and a regional mechanism is proposed that explains both the observed decline, and the partial recovery since the late 2000s.

RESULTS
Figure 1a shows the decline in Eastern Africa MAM rainfall during 1998-2008 relative to 1986-1997, and recent recovery from 2010 onwards, although with high year to year variability.The timeseries is divided into three periods: a wetter period (1986-1997, P1), a drier period (1998-2008, P2) and the recent partial recovery (2009-2018, P3).Analysis of the seasonal cycle (Fig. 1b) reveals that the decrease in rainfall during P2 is clearly associated with a delayed onset (Fig. 1c, 4.9 days on average) and earlier cessation (Fig. 1d, 2.2 days on average), rather than a change in seasonal maximum daily rainfall as often assumed.Onset in P3 recovers to P1 values (Fig. 1e), while cessation in P3 partially recovers but is still earlier than during P1 (Fig. 1f).Supplementary Fig. 2 shows the onset and cessation dates for individual years.These observed changes in seasonality are not fully explained by the warming of the Indo-Western Pacific SSTs altering subsidence over Eastern Africa since there is no decrease in seasonal maximum in areaaveraged daily rainfall during P2 (Fig. 1b).
A delayed and then faster movement of the tropical rain-band northwards across Eastern Africa during the boreal spring could explain later onset and earlier cessation (Supplementary Fig. 3).Previous studies, using multiple observational datasets, found wetting at the northern and southern limits of the tropical rain belt extent during June-August and December-February over A recent study highlighted that under future climate change the strengthened SHL delays the southward progression of the rainband across Africa, giving a later cessation and onset 13 .The recent observed faster northward movement of the rain-band in MAM therefore similarly suggests the SHL or a comparable mechanism may play a role in the earlier long-rains cessation revealed in Fig. 1.Following this study 13 , here changes in a different, proximal, heat low are investigated, to ascertain whether this influences the recent changes in the timing of the long rains over Eastern Africa.

Early cessation
Focusing on the early cessation, the Arabian Heat Low is a dominant feature over the Arabian Peninsula in spring and summer 16 , which interacts with the Somali Jet and Indian Monsoon [17][18][19] .Figure 2 (and Supplementary Fig. 6) reveals a strong decrease in geopotential height over Arabia and the adjacent Arabian Sea in May in ERA-Interim from P1 to P2 (−8.46 m over 50-75°E, 10-25°N).This is also present in April (further north), is seen in other reanalyses (NCEP and MERRA) and is robust to changing periods of the composite (for P3-P1 see Supplementary Fig. 5 and Supplementary Fig. 6).The decrease in geopotential height is expected to draw the rain-band northwards faster and further, consistent with the observed and analysed rainfall changes (Figs. 1, 2, Supplementary Fig. 4).ERA-Interim captures the drying in Central and Eastern Africa in May, and the northward shift in Indian Ocean and Western India rainfall seen in observations 20 (Supplementary Fig. 4).The long rains correspond to the seasonal northerly-to-southerly reversal of the Somali Jet off the Horn of Africa, as the tropical rain-band crosses Eastern Africa.During P2 the southerlies are enhanced across Eastern Africa in May (Fig. 2), consistent with the earlier cessation of the faster propagating rain-band.Temperatures over the Arabian Peninsula increased over 1979-2009 21 , consistent with a pressure and geopotential height decrease.Furthermore, Arabian July-September temperatures exhibited a somewhat step-wise increase in the mid-1990s 22 , coincident with the Eastern African rainfall decline.
The decline in Eastern African May rainfall from 1998/9 onwards (coincident with the decrease in MAM rainfall) followed the decrease in geopotential height over Arabia starting in 1997 (Fig. 3).The correlation between the geopotential height and May rainfall (Climate Hazards Group infra-red Precipitation with Stations (CHIRPS)) time-series is moderate (r = 0.59), but significant at the 1% significance level (Fig. 3a, b).This correlation increases with smoothing, which suggests that both are driven, at least to some extent, by a common relatively persistent driver, with SSTs the most obvious candidate.This relationship might also address some of the low predictability of MAM rainfall on a seasonal time-scale 23,24 .
The AMIP multi-model mean captures the drying across Eastern Africa in March and May and the increase in the strength of the Somali Jet in May (Fig. 2).It also captures the decrease in geopotential height in May over the north Arabian Sea (Fig. 2f), although the decrease is further southeast than in ERA-Interim, and restricted to over the ocean, not extending sufficiently over Arabia (Fig. 2 and Supplementary Fig. 6).The decrease in precipitation and geopotential height are of lower magnitude than the observed changes which may be linked to the lack of drying across the CMIP5 simulations.The assimilation of observations into ERA-Interim helps capture the trends compared with AMIP.AMIP simulations struggle to fully capture the response over land to SST and other forcings 25,26 .Summertime dust increases over Arabia from 1997 to 2009 have been shown to deepen the Arabian Low in simulations 26 , suggesting that the dust trend may influence Eastern African MAM rainfall.
We correlate the change in May geopotential height southeast of Arabia (where the AMIP models give the largest pressure decrease and ERA-Interim also shows a decrease) with both the change in May rainfall and the change in long rains cessation over Eastern Africa across the AMIP simulations (Fig. 3 and Supplementary Fig. 7).The correlation (r = 0.33) shows that the spread in the geopotential height decrease between models explains some of their range in the change in May rainfall and cessation date of the long rains.The multi-model mean captures the interannual correlation between the geopotential height over Arabia and the northern Arabian Sea and Eastern African May rainfall (Supplementary Fig. 8), with 25 out of 28 models capturing the positive From P1 to P2, May SST increased more over the very north Arabian Sea than further south in the Indian Ocean (Fig. 4b).Warm May SSTs in the Indian Ocean are associated with a later cessation (Fig. 4d), but this correlation is reversed over the very north Arabian Sea (north of 15-20°N).We suggest that these changes in SST gradient across the Arabian Sea are associated with both the rainfall variability and its trend, with warmer SSTs to the north and cooler SSTs to the south initiating a pressure gradient across the Arabian Sea, drawing the rain-band north.Also, warmer SSTs to the north will favour moist convection there.The withdrawal of rainfall from Eastern Africa at the end of the long rains has previously been associated with the establishment of a lowpressure zone and active convection in the Arabian Sea 27 .Two regions 28 were chosen to calculate the north-south SST gradient across the Arabian Sea (Fig. 4f), which correlates with the longrains cessation at −0.61.The correlation was tested for robustness using 10 SST realisations (see methods) and all correlations were found to be significant, with a range of −0.45 to −0.70 across the 10 realisations.This supports the hypothesis that the rain-band has been drawn north by the warmer SSTs and the associated lowpressure, with an earlier cessation linked with relatively warmer SSTs to the north from P1 to P2. Comparing P3 with P2 (Supplementary Fig. 9), the SST has warmed more over the very north Arabian Sea, but contrast in warming is less, contributing to the partial recovery in long rains cessation, amongst other factors.

Late onset
Turning to the late onset, warming SSTs south of Madagascar (Fig. 4a) are associated with reduced March rainfall and later onsets (Fig. 4c), as warmer SSTs to the south delay the northward progression of the rain-band.Figure 4e 43 (see methods).Stippling indicates a p value less than 0.1.e, f Scatter plot of the SST gradient in March/May (green box minus purple box) with the mean onset/cessation date from CHIRPS across East Africa for each year.The values quoted in the title are the r and p values calculated using Pearson correlation coefficient long rains onset and the southern-hemisphere SST gradient.From P2 to P3 March SST cooled south of Madagascar (Supplementary Fig. 9), potentially explaining the recent recovery in onset.Furthermore, during P2 the northerlies are enhanced in March (Fig. 2), also consistent with the delayed onset.

Future climate change
Sub-tropical land is expected to warm faster than tropical oceans under climate change 29 , which together with water-vapour feedbacks can deepen sub-tropical heat lows, affecting rainband progression [13][14][15]30,31 . CMIP5projections for 2080-2100 suggest a greater and more extensive deepening of the SHL compared to the Angola Low 13 .This leads to a strengthening north-south asymmetry with the rain-band moving further north in boreal summer, but not further south in austral summer.CMIP5 projections also show earlier onset and earlier cessation for Eastern Africa in MAM 13 .Considering the regional mechanisms identified here, projections under RCP 8.5 show a greater increase in geopotential height over the Indian Ocean compared with Arabia in May and strengthening of the Somali Jet (Supplementary Fig. 10), consistent with projections of an earlier cessation of the long rains 13 .In March, the surface temperature increases more in the western Indian Ocean than south of Madagascar, which combined with the southerly wind anomaly is consistent with projections of earlier onset 13 .Thus, CMIP5 projections of earlier onset and cessation are consistent with the mechanism we have identified to describe the recent decline. Th overall increase in long-rains rainfall projected under future climate change 32 (that is less than the projected increase in the short rains) 6 , is therefore the consequence of little change in overall season length 13 and an increase in the intensity of rainfall on individual days 13,32,33 .

DISCUSSION
In summary, we find that the observed decline in Eastern African Long Rains is characterised by a shortening of the rainy season (with later onset and earlier cessation) rather than by a decrease in the peak daily rainfall.The cause of the shortening is a faster movement of the rain-band over Eastern Africa during the boreal spring.This is a consequence of warmer SSTs to the north during boreal summer and to the south during austral summer, which leads to an increased pressure gradient and hence to more rapid travel of the rain-band during the whole January-August period.The hypothesis that reduced pressure over Arabia and the adjacent ocean influences rain-band progression is supported by the fact that this pressure change explains some of the variation in May rain change across the AMIP ensemble.The results highlight the interhemispheric transitional nature of the long rains, with the locations of March and May drivers in opposite hemispheres.Further studies should quantify to what extent the mechanisms for rainfall decline revealed in this study are driven by natural variability, amplified land-ocean pressure gradients driven by anthropogenic carbon emissions and changing aerosol forcings; and the role of these mechanisms under future climate change.

Precipitation data
The CHIRPS daily rainfall dataset 34 and Tropical Applications of Meteorology using Satellite data and ground-based observations (TAMSAT) 35 daily rainfall dataset were used for 1985-2018.Both rainfall datasets use thermal infra-red imagery to calculate rainfall totals; CHIRPS also includes gauge data, a monthly precipitation climatology, and atmospheric model rainfall fields from the NOAA Climate Forecast System, version 2 (CFSv2) 34 .TAMSAT, however, just uses thermal infra-red imagery; gauge observations are used for the time-invariant calibration but are not incorporated into the estimates 35 .The Global Precipitation Climatology Project (GPCP) monthly precipitation analysis was also used over 1985-2018; it uses low-orbit satellite microwave data, geosynchronous-orbit satellite infra-red data, and surface rain gauge observations to calculate rainfall totals 36 .

Reanalysis data
Geopotential height and wind data were taken from the ERA-Interim reanalysis for 1986-2018, produced by the European Centre for Medium-Range Weather Forecasts (ECMWF), using the Integrated Forecast System combined with data assimilation 37 .

SST data and correlations
Monthly HadISST (v1.1) observed SST data, produced by the UK Met Office were used for 1986-2018 38 .In order to compute the uncertainty on the SST correlations, ten different realisations of HadISST (v2.2) were used to compute a range on the correlation values 39 .These data are only available until 2015, and hence are used for 1986-2015.Correlations were computed using the Pearson Correlation Coefficient, and data were detrended first.

Atmosphere-only model simulations
Atmosphere-only simulations were obtained from 28 models from the CMIP5 generation of models 40 .The atmosphere-only simulations are driven by historical SSTs, sea ice and radiative forcing agents.Daily and monthly precipitation and monthly geopotential height and winds at 850 and 925 hPa were obtained for 1986-2008.Only the first ensemble members (r1i1p1) are used.To enable the construction of multi-model means, the data were re-gridded to a 1°× 1°grid.The models used are listed in Table 1.

Coupled model simulations
Data from 29 coupled climate model simulations, used in CMIP5 40 , were used to assess changes in variables under future climate change.The coupled simulations include a fully coupled ocean and are driven by historical radiative forcings for the observation period and use radiative forcings from representative concentration pathway (RCP) 8.5 for the future projections 41 ; this is a high emission scenario, with a radiative forcing of 8.5 W m −2 at 2100.Only the first ensemble members (r1i1p1) are used.Monthly geopotential height (850 hPa), near-surface air temperature and eastward and northward winds (850 hPa) were obtained for 1980-1999 and 2080-2099.To enable the construction of multi-model means, the data were re-gridded to a 1°× 1°grid for geopotential and temperature, and 3°× 3°for winds.The models used are listed in Table 2.

Onset/cessation methodology
Onset and cessation dates were computed using the methodology of anomalous accumulation, used in a number of studies of seasonality over Africa 42,43 .This methodology identifies wet seasons when the rainfall is persistent in occurrence, duration and intensity 44 .Full details of the methodology can be found in Dunning et al. 43 ; the two season variant is used here.Firstly, the two periods of the year when the wet seasons occur, labelled the climatological wet seasons, are determined by identifying local minima and maxima in the climatological cumulative daily mean rainfall anomaly.To calculate the climatological cumulative daily mean rainfall anomaly, the climatological mean rainfall for each day of the calendar year, R i , and the long-term climatological daily mean rainfall, R, are computed.The climatological cumulative daily mean rainfall anomaly on day d, is where i ranges from 1st January to the day for which the calculation applies.Local minima and maxima in C(d) determine the beginning and end of the climatological seasons; if the method cannot identify two seasons then the point is excluded from this analysis, as the main interest is the biannual regime over East Africa.Onset and cessation dates are calculated for each season individually, by calculating the daily cumulative rainfall anomaly for each climatological wet season (starting 20 days prior to the start of the climatological wet season and ending 20 days after the end of the climatological wet season).The minima in the daily cumulative rainfall anomaly is the onset date, and the maxima is the cessation.Seasons of less than 14 days in length are excluded.

Regions used in the analysis
Harmonic analysis was used to determine the region of East Africa that experiences a biannual regime.Using the CHIRPS daily rainfall data, the amplitude of the first harmonic and second harmonic were computed at each grid point.The ratio was then calculated; if the ratio was greater than 1, i.e., the amplitude of the second harmonic was greater than the amplitude of the first harmonic, then the grid point was defined as biannual.The biannual region was then smoothed using a Gaussian filter to give the region mask shown in Fig. 1c-f.For other datasets (e.g., AMIP data) the same mask was re-gridded.
For the correlations in Fig. 3, the Arabian Heat Low was defined to be 50-75°E, 10-25°N; this region captures the decline in geopotential in both ERA-Interim and the AMIP multi-model mean.

Fig. 1
Fig. 1 Changes in seasonality of the long rains.a MAM rainfall anomaly over Eastern Africa (see region in c-f) 1985-2018 from CHIRPS, TAMSAT and GPCP.The bars represent the anomaly for each year (CHIRPS); the lines are smoothed using a 3 year moving average.b Mean seasonal cycle over Eastern Africa from CHIRPS over the 3 periods of study.c, d Maps showing the mean change in onset (c) and cessation (d) date of the long rains from 1986-1997 to 1998-2008.Onset and cessation dates were calculated using CHIRPS data and the method of Dunning et al. 43 (see methods).The purple contour demarcates the region used for Eastern Africa.e, f, As (c, d), but comparing 2009-2018 with 1986-1997

Fig. 2
Fig. 2 Change in precipitation, geopotential height and 850 hPa winds for P2 minus P1.Change in precipitation (coloured shading), 850hPa geopotential height (m, green contours: solid lines represent positive values and dashed lines represent negative values) and 850 hPa winds (black arrows), for the period P2 minus period P1, for March (a, b), April (c, d) and May (e, f).Data for (a, c, e) are from ERA-Interim (wind and geopotential height) and GPCP (rainfall) and for (b, d, f) are from the AMIP multi-model mean (all variables)

Fig. 3
Fig. 3 Relationship between changes in geopotential over the north Arabian Sea and changes in East African May rainfall.a, b Mean May 850 hPa geopotential height over the North Arabian Sea from ERA-Interim and mean May precipitation over East Africa from GPCP and CHIRPS over 1986-2018 smoothed with a 3 year moving average.The corresponding scatter plot (CHIRPS) is shown in (b).Correlation was calculated on the unsmoothed time-series.c The median change in cessation date of the long rains over East Africa from P1 to P2 in each AMIP model is plotted against the mean change in 850 hPa geopotential height over the north Arabian Sea from P1 to P2.The numbers correspond to the numbers for each AMIP model in Table 1 (methods); the cross shows the multi-model mean.The circle shows the observations; the change in geopotential height was calculated from ERA-Interim, and the change in cessation from the cessation dates computed using CHIRPS data shows a correlation of −0.41 (−0.29 to −0.48 across 10 realisations of HadISST) between

Fig. 4
Fig. 4 Changes in March/May SST and relationship with Long Rains onset/cessation.a, b Change in March (a) and May (b) SST from P1 to P2 from HadISST.c, d Correlation (Pearson correlation coefficient) of onset (c) and cessation (d) date for the long rains across East Africa with March (c) and May (d) SST over 1986-2018.Onset and cessation dates were calculated using CHIRPS data and the method of Dunning et al. 43 (see methods).Stippling indicates a p value less than 0.1.e, f Scatter plot of the SST gradient in March/May (green box minus purple box) with the mean onset/cessation date from CHIRPS across East Africa for each year.The values quoted in the title are the r and p values calculated using Pearson correlation coefficient

Table 2 .
List of CMIP5 models and institutions that provided coupled model output used in this study

Table 1 .
List of AMIP models (and institutions) used in this study.The number in the first column is the symbol used for that model in Fig.3c 64Published in partnership with CECCR at King Abdulaziz University npj Climate and Atmospheric Science (2019) 34