1 Introduction

Climate variability influences all facets of human activity, particularly in agrarian economies. To ensure the sustainable use of resources, it is appropriate to develop climate predictions based on scientific understanding that is tailored to specific areas and impacts (Glantz et al. 1991; Glantz 1996; Moura and Sarachik 1997; WCRP 1999; Hammer et al. 2000; Jury 2002). The predictability of climate derives primarily from slowly evolving ocean conditions. Using sea surface temperature (SST) forecasts in coupled general circulation models, a measure of success has been achieved in forecasting African rainfall at one season lead time (Lyon and Mason 2009). Atmospheric variables may be skilful in climate prediction in the context of wind forced ocean Rossby waves and equatorial upwelling dynamics (Jury et al. 1997; Xie et al. 2002; Jury and Huang 2004).

The interaction of SST and atmospheric convection associated with the near-equatorial trough is an important driver of tropical climate variability. As the surface layer of the equatorial ocean cools under the influence of wind-driven upwelling and surface heat loss, atmospheric convection shifts. This alters the global circulation and hydrological budget over Africa (Lamb 1978, 1982; Nicholson 1981, 1998; Nicholson and Palao 1993; Nicholson and Grist 2001). Past studies have attributed African climate variability to meridional gradients of Atlantic SST (Folland et al. 1986; Lough 1986; Palmer 1986; Semazzi et al. 1996) and to the Pacific El Niño-Southern Oscillation (ENSO; Fontaine et al. 1995; 1998). During ENSO warm phase upper westerly winds prevail over the Central Atlantic. Downstream over Africa, these winds descend, weakening the Hadley circulation (Hastenrath 2000; Camberlin et al. 2001). In addition to ocean-atmosphere coupling, land-atmosphere feedback on the boundary layer climate provides a forcing that has received attention (Charney 1975; Xue et al. 1990; Zeng and Eltahir 1998; Zeng et al. 1999).

While a majority of studies have concentrated on Africa’s Sahel climate west of 20ºE, here, the focus is directed toward understanding decadal climate variability of the Ethiopian highlands, which impacts on the flow of the Nile River (Quinn 1992; Eltahir and Wang 1999; Wang and Eltahir 1999). Ethiopian climate variability has been related to ENSO and the Southwest monsoon over the Arabian Sea (Camberlin 1995, 1997; Camberlin et al. 2001). Deep convection over India establishes a temperature gradient that drives a tropical upper level easterly jet over the Arabian Sea (Segele et al. 2009). When the jet surges, boreal summer rainfall over Ethiopia tends to increase. Ethiopian rainfall is suppressed following an active tropical cyclone season in the Southwest Indian Ocean (Shanko and Camberlin 1998).

The objective of this paper is to understand the large-scale atmospheric circulations that modulate Ethiopian climate at decadal time scales during the twentieth century. The forcing mechanisms are studied through analysis of historical datasets and statistical analyses. Another goal of this study is to understand how the Walker and Hadley circulations interact over Africa, extending the earlier work of Yeshanew and Jury (2007). A number of scientific questions are addressed: (1) What are the characteristics of Ethiopia’s decadal climate variability? (2) Which environmental fields are most important and how do they evolve? And (3) What role does the ocean play in the Walker–Hadley interaction? In Section 2, the data and methods are outlined. Section 3 provides results by analysis of mean climate, rainfall and vegetation modes, large-scale environmental factors, and composites for wet and dry summers. The conclusions are given in Section 4 within a conceptual framework.

2 Data and methods

To study the climate of Ethiopia (3–17° N, 33–47° E, Fig. 1a) monthly gauge rainfall from the World Climate Research Program Global Precipitation Climatology Center (GPCC), dataset is extracted from the International Research Institute’s Climate Library over the period 1901–2007. Although GPCC rainfall records are quite complete and quality checked, data density in the period prior to 1950 is somewhat limited across Ethiopia (Conway et al. 2004, Fig. 1c). The GPCC monthly rainfall data are optimally interpolated to a 0.5° grid as described in Rudolf and Schneider (2005) to account for orographic effects. Here, the seasonal cycle is filtered out, and normalized rainfall anomalies are linearly detrended and smoothed with a centered 60-month running mean to eliminate high-frequency fluctuations and drift. The filter suppresses 3- and 6-year cycles which have been previously studied (Yeshanew and Jury 2007). To investigate the nature of climatic oscillations, the filtered rainfall time series are subjected to wavelet spectral analysis. The continuous 5-year filtered rainfall fields are objectively clustered using singular value decomposition (SVD) as described by Bretherton et al. (1992); von Storch and Navarra (1995); and Venegas (2001). SVD analysis involves an eigenvector decomposition of the covariance matrix within a single input field, such that the variability is reduced to modes of decreasing normalized variance. For each mode, there is a spatial pattern of loadings that describe its area of influence and time scores that reveal the amplitude and wavelength of oscillation. Varimax rotation is imposed to limit the effects of domain size.

Fig. 1
figure 1

1950–2000 data density for a SST measurements per month per 2º box and b GPCC rain gauges per 2º box. c Time series of GPCC raingauge density in Ethiopian highlands area (box in b)

Full size image

In a similar manner, SVD is applied to monthly reanalyzed atmospheric fields (Kalnay et al. 1996) of National Center for Environmental Prediction (NCEP) zonal and meridional winds, and stream function in vertical sections averaged over the African hemisphere, and to NOAA-extended SST (Smith et al. 2008) with the same filtering, over the 1949–2007 time period. SVD is also applied to bimonthly satellite vegetation (NDVI) data in the period 1981–2004 to explore how the annual cycle affects the interaction between Ethiopian climate and surface vegetation, with filtering limited to a 3-month running mean (seasonal cycle is retained). To quantify the influence of (mainly NCEP) environmental variables on filtered Ethiopian rainfall, SVD time scores are used in lag cross-correlation. Correlation values >0.40 are considered significant at the 95% confidence limit for a degrees of freedom of ∼22 (in the period 1949–2007) that takes into account the reduction of sample size by the 5-year filter.

Composite environmental conditions are evaluated on a seasonal basis, by ranking the top 10 wet and dry summers (May–October) based on GPCC data averaged over the Ethiopian highlands (4–14° N, 34–42° E). Wet summers include: 1959, 1962, 1963, 1965, 1966, 1979, 1998, 1999, 2006, and 2007. Dry summers are 1971, 1972, 1973, 1984, 1985, 1986, 1992, 2002, 2003, and 2004. The global structure of GPCC rainfall, NOAA-SST, and NCEP upper winds are analyzed by subtracting dry seasons from wet. This helps place Ethiopian decadal climate variability in a wider context. The composite structure of humidity, vertical motion, and zonal and meridional overturning circulations is studied as vertical sections over Africa.

Figure 1a–c addresses the issue of data density. The map of ship-based SST measurements in the period since 1950 indicates that the North Atlantic and main shipping lanes are well covered. Elsewhere, less data are available. The rain gauge network (Fig. 2b) is adequate for the study of decadal climate variability across most land areas, and Ethiopia appears to be no worse than more developed regions. Yet, coverage is limited prior to 1950 (Fig. 1c). To overcome this problem, modern satellite patterns are projected onto historical data to create realistic patterns, using optimal interpolation methods similar to SVD. In this way, the Ethiopian climate can be linked in time and space with surrounding information to create a more complete coverage that can be analyzed as described here. SVD has the value of picking up more recent patterns when the data density is high and projecting them back in time over the period of lower data density. The 5-year filter also helps close data gaps without the need to artificially supplement the record.

Fig. 2
figure 2

a Topographic map of Africa with Ethiopia outlined, b mean 850 hPa winds in July and country names, c mean GPCC July rainfall in millimeter per month, d mean July vegetation fraction from satellite NDVI with dashed line for Rift Valley, e mean NCEP 500 hPa vertical motion (Pa/s) and winds in July, f Hovmoller time-longitude plot of mean NCEP 700 hPa relative humidity on 9° N, describing seasonal cycle

Full size image

3 Results

3.1 Modes of Ethiopian climate

Ethiopia is elevated ∼2,000 m above the surrounding areas (Fig. 2a), and orographic forcing of convection is common (Fig. 2c). During summer, the southwest monsoon blows along the coast of Somalia (Fig. 2b). It instills cyclonic vorticity and helps draw tropical air from the Congo Basin to the Ethiopian highlands, where vegetation cover is dense (Fig. 2d). Over Southern Sudan, there is a region of rising motion at 500 hPa that extends into western Ethiopia (Fig. 2e) along the axis of easterly flow. Although moisture is present over Ethiopia through most of the year (Fig. 2f), it is most zonally extensive in midsummer (July) when a zonal trough lies across the region. During this season, the Indian Monsoon generates an upper easterly jet that triggers pulses of rainfall. Easterly waves are initiated over the Ethiopian highlands from May to October (Lin et al. 2005) and propagate across the African Sahel into the tropical Atlantic.

The leading SVD modes for vegetation in the period 1981–2004 are illustrated in Fig. 3a–c, with seasonal cycle and high-frequency variability retained. Mode 1 is represented as a cluster over Northwestern Ethiopia that extends into Sudan. Its time score has a 12-month cycle, indicating that rainfall from vortices in the near-equatorial trough reach this latitude by midsummer and retreat thereafter. Vegetation in Northwestern Ethiopia is least productive in February–March and gradually rises to a peak in August–September. Mode 2 clusters over the southern half of Ethiopia and extends into Northern Kenya. It exhibits a 6-month cycle related to early and late summer rainfall maxima inducing two distinct seasons of vegetation growth, typical for near-equatorial zones. Inspection of the mean time score error bars representing the standard deviation (Fig. 3c) reveals that the early summer vegetation peak (April–May) is less variable than late summer (October–November). The vegetation SVD analysis divides Ethiopia into two climatic regimes: northwestern unimodal and southern bimodal. The question is whether climate variability should also separate along these lines.

Fig. 3
figure 3

a SVD loading pattern for mode-1 satellite vegetation, b SVD loading pattern for mode-2 satellite vegetation, and c mean mode-1 (inverted) and mode-2 time scores and standard deviations (error bars) contrasting the annual cycle. Dashed line in (a) is Rift Valley. Loading pattern color bar has units of normalized anomalies from −3 to +3

Full size image

GPCC rainfall modes in the period 1901–2007 (Fig. 4a, d) feature a cluster in the north as mode 1 and a cluster in the south as mode 2, representing 19% and 13% of total variance, respectively. Hence, the vegetation modes correspond with rainfall despite differences in filtering and time period. The northern mode-1 cluster extends westward into Sudan. The southern mode-2 cluster reaches eastward into Somalia. Time scores for the filtered rainfall SVD modes are given in Fig. 4b, e and wavelet spectral analysis highlights the amplitude and wavelength in Fig. 4c, f. Northern Ethiopian rainfall displays a persistent decadal cycle throughout the twentieth century, as found earlier (Wood and Lovett 1984; Seleshi et al. 1994). Southern Ethiopian rainfall, on the other hand, has a weak 10–12-year cycle early in the record, and a 14–18-year cycle in the period 1950–1990. Inspection of spectral energy in unfiltered rainfall records suggests that the decadal oscillations are a real feature and not a statistical artifact. When area averages of 5-year smoothed rainfall in the two regions are analyzed, their time series are uncorrelated (r = −0.10). This is seen as a strategic advantage for Ethiopia: different parts of the same country have independent climatic regimes.

Fig. 4
figure 4

a SVD loading pattern for mode-1 North Ethiopia 5-year filtered rainfall anomalies, b 5-year filtered temporal variability, and c wavelet spectra with cone of validity; d loading pattern for mode-2 South Ethiopia rainfall, e 5-year filtered temporal variability, and f wavelet spectra with cone of validity. Color bar for loading patterns as in Fig. 3. Units of b, e are converted to millimeters per month

Full size image

3.2 Environmental links

Mapping the correlation of global climate indices onto gridded GPCC summer (May to October) rainfall time series, the opposing character of rainfall regimes in northern and southern Ethiopia emerges. Of a possible 30 indices available from the NOAA Climate Diagnostics Center, three achieve robust associations in the period 1949–2007 (Fig. 5a–c). Pacific SST Nino3 correlations are significantly negative northwest of the Rift Valley and positive to the southeast. Hence, ENSO warm phase may be associated with a reduced northward migration of the ITCZ, such that rainfall is confined near the equator. Pacific SSTs are also known to modulate the zonal Walker Circulation over the Indian Ocean, which in turn affects the end of the rainy season in southern Ethiopia. The Pacific Decadal Oscillation (PDO, Tourre et al. 1999; Barlow et al. 2001) also exhibits a strong negative correlation northwest of the Rift Valley and weak positive association to the southeast. The Atlantic Multidecadal Oscillation (AMO, Enfield et al. 2001) correlation pattern is reversed: warm phase enhances rainfall in northwest districts and suppresses to the southeast.

Fig. 5
figure 5

Zero-lag correlation of unfiltered summer GPCC gauge rainfall 1949–2007 with various climatic indices: a Nino3, b Pacific decadal oscillation, and c Atlantic multi-decadal oscillation

Full size image

The zonal and meridional overturning atmospheric circulations are characterized over the African hemisphere by application of SVD to NCEP U and V wind components and stream function in vertical sections from 1,000–100 hPa averaged over the latitudes 10° N–5° S for east–west sections and longitudes 60° W to 90° E for north–south sections. In addition, SVD is applied to NOAA SST fields, with consistent filtering (e.g., detrended anomalies with 5-year running mean). Patterns that correspond with Ethiopian rainfall variability are shown in Fig. 5a–h, while cross-correlations are listed in Table 1.

Table 1 Cross-correlation of continuous 5-year filtered Ethiopian rainfall with filtered environmental SVD time scores in the period 1949–2007
Full size table

Ethiopian rainfall modes 1 and 2 relate to north–south sections of stream function and meridional wind, respectively. The stream function (psi) SVD loading pattern is tilted such that the low level cyclonic maximum underlies the upper level anticyclonic maximum, reflecting a Walker Circulation across the African hemisphere (Fig. 6a). The time scores correspond (r = −0.57, Fig. 6e) and lag-correlations remain significant to 36-month lead time (Fig. 6g). This linkage is consistent with SST mode 1 (Fig. 6c) that reflects the AMO (r = −0.83). Southern Ethiopian rainfall mode 2 is well correlated with meridional wind mode 3 (Fig. 6b, Table 1). The filtered SVD time scores correspond (r = −0.70) with lag-correlations significant to 24-month lead time (Fig. 6h). The association is modulated by SST mode 2 (Fig. 6d) that reflects the PDO (r = −0.60). The SVD time score correlations (Table 1) suggest that the Hadley circulation helps activate rainfall over southern Ethiopia. Although correlation values are inflated by the 5-year filter, the structure of the lag-correlation function is informative. It suggests multiyear predictive potential from atmospheric derivatives that characterize the global overturning circulations. SST modes and indices, on the other hand, seem to lag the overturning atmospheric circulations in the decadal band.

Fig. 6
figure 6

SVD loading patterns for a vertical N–S section mode-1 stream function (PSI-1) and b mode-3 meridional wind (V-3); SVD loading maps for c mode-1 SST and d mode-2 SST; 5-year filtered time score comparison between e Northern and f Southern Ethiopia rainfall and leading environmental modes; lag-correlation plot between g PSI-1 and Northern Ethiopia rainfall and h V-3 and Southern Ethiopia rainfall. Positive values on X-axis (g, h) refer to variable leading rainfall. Environmental variable refers to left-hand inverted Y-axis in (e, f)

Full size image

3.3 Composite differences

While the continuous 5-year filtered SVD rainfall analysis has revealed certain aspects of Ethiopia’s decadal climate variability, it is useful to place the work into a wider context using seasonal data. In this regard, composites were constructed by averaging fields for the top 10 dry and wet summers (May–October) since 1949 and subtracting to obtain the difference. Results are given in Fig. 7a–c for rainfall, SST, and upper winds. Rainfall is found to increase >30 mm/month across the 0–20º N zone influenced by the near-equatorial trough, from Central America eastward across the African Sahel and Southern India to Southeast Asia. SST differences are below normal across the southern subtropics in the vicinity of the semi-permanent anticyclones. The extra-tropical Pacific exhibits negative SST differences in a PDO pattern, while the North Atlantic has positive differences consistent with AMO warm phase. The upper wind field in respect of Ethiopian wet minus dry conditions is strongly easterly (10 m s−1) over the Central Atlantic and consistent with Walker circulation adjustment to ENSO/PDO cool phase (Yeshanew and Jury 2007).

Fig. 7
figure 7

Composite wet minus dry differences in respect of Ethiopian rainfall for a gauge rainfall (mm/month), b SST (ºC), and c 150 hPa wind (m/s). Dashed box in (a) highlights zone of anomalous rainfall and near-equatorial trough

Full size image

Vertical north–south sections are constructed over African longitudes for the wet minus dry summer (May–October) composite and illustrated in Fig. 8. Moisture is enhanced in the 1,000–500 hPa layer from 15–20° N (Fig. 8a) and suggests a northward displacement of the near-equatorial trough in conjunction with rising motion over the northern Sahel (Fig. 8b). Spin-up of the southern Hadley cell is evident in meridional wind differences over the equator (Fig. 8c): southerly in the low levels (+V, 800–600 hPa) and northerly in the upper levels (−V, 200–150 hPa). The anomalous southern Hadley circulation coincides with upper easterly (–U)/lower westerly (+U) wind differences associated with the Atlantic Walker cell (Fig. 8d). Analyzing differences in the zonal and meridional overturning circulations, it is found that they are well established 2 years before anomalous summer rains are observed over the Ethiopian highlands (Fig. 8e, f). This is not an artifact of filtering but a reflection of the slow rhythm in overturning circulations that modulate decadal climate variability.

Fig. 8
figure 8

Vertical north–south section composite wet minus dry differences in respect of Ethiopian rainfall for: a specific humidity (g/kg), b omega (Pa/s), c meridional wind (m/s) representing southern Hadley cell, d zonal wind (m/s) representing Atlantic Walker cell, e meridional wind 2 years prior, f zonal wind 2 years prior, all averaged over African longitudes 20° W to 50° E

Full size image

While the composites have indicated how the atmospheric circulations relate to tropical rainfall, an objective SVD analysis of zonal and meridional wind sections reveals how the low frequency component of the Walker and Hadley cells interact over Africa (Fig. 9). The mode-1 Walker circulation across the Atlantic corresponds closely with the mode-1 Hadley circulation over Africa (r = −0.85). When upper westerlies/lower easterlies are established, meridional convergence in the lower branch of the Hadley cell is suppressed. The time scores (Fig. 9c) indicate the Hadley cell leads by about 1 year in the 1950–1980 period and lags by a few months thereafter. Although Walker cells tend to alternate faster than the Hadley cell, their decadal resonant frequency is consistent with anomalous north–south displacements of the near-equatorial trough and wet–dry cycles of Ethiopian rainfall.

Fig. 9
figure 9

Vertical sections of SVD loading pattern for mode-1 a zonal wind (U1w) across longitudes averaged 5° S–10° N, b meridional wind across latitudes (V1h) averaged 60° W–90° E, and c their time scores with V1h-inverted. African highlands are represented by shaded rectangle

Full size image

4 Conclusions

Ethiopian decadal climate variability has been characterized in this study. There are two distinct rainfall modes that correspond with unique seasonal cycles and opposing responses to regional and global forcing. The northern zone that impacts the Nile River and underlies the tropical easterly jet has a unimodal rainy season that is enhanced by an AMO SST pattern in warm phase. This northern mode links with the Atlantic Walker zonal overturning circulation (PSI-1) and has exhibited 10–12-year cycles through much of the twentieth century. The southern zone has a bimodal rainy season (mode 2) that is enhanced by a PDO SST pattern in cool phase and the southern Hadley meridional overturning circulation. Multiyear wet and dry spells were uncovered in Ethiopian rainfall records. The wet periods were: 1962–1966, 1979, 1998–1999, and 2006–2007; dry periods included 1971–1973, 1984–1986, 1992, and 2002–2004. Decadal climate oscillations require attention because of their effect on food and water resources. It is possible for agrarian communities to set aside resources for one “bad” year, but more than two in a row leads to socioeconomic disruption. Prediction of the amplitude and timing of multiyear wet and dry spells is possible given that the Walker and Hadley signals tend to lead SST and rainfall by over 1 year (Fig. 6). Further efforts to better understand and predict the variable climate of Ethiopia will involve relating intra-seasonal weather events to the background climate and determining how localized land-atmosphere feedbacks conspire with remote teleconnections.