1 Introduction

Teleconnections are recurring and persistent, large‐scale patterns of pressure and circulation anomalies that span vast geographical areas (http://www.cpc.ncep.noaa.gov/data/teledoc/teleintro.shtml). They are preferred modes of low‐frequency (long time scale) variability, such as quasi‐global oscillations [the El Niño Southern Oscillation (ENSO), North Atlantic Oscillation, etc.]. Commonly, teleconnections are identified and analysed using principal component analysis (PCA) or empirical orthogonal functions (EOF); see among others Halldor and Venegas (1997). Recently however, the empirical orthogonal teleconnection (EOT) analysis was introduced by van den Dool et al. (2000) as an alternative method to find large‐scale atmospheric patterns (Simonti and Eastman, 2010). In earth observation time series data, EOTs examine each pixel over time to find which pixel can explain the greatest amount of variability of all other pixels combined (Clark Labs, 2009). As shown by van den Dool et al. (2000), EOT patterns are very similar to corresponding rotated EOFs, but computation is far more efficient.

Quasi‐global teleconnections are important for explaining drought occurrence. Of all natural hazards, droughts may be the most complex and least understood phenomena (Modarres and Ouarda, 2014), partly because of strong associations with quasi‐global oscillations (Hoerling and Kumar, 2003; Behera Swadhin et al., 2005). For example, the ENSO is known to influence drought occurrence in, among others, Brazil (Liu and Juarez, 2001), Indonesia, the Philippines (Harger, 1995) and Ethiopia (Abtew et al., 2009).

North Ethiopia was chosen as a case‐study region for testing the effectiveness of the EOT method to identify the impact of quasi‐global oscillations on drought risk (Figure 1(a)).

Earlier studies have linked (North‐) Ethiopian rainfall with atmospheric teleconnections. Among the most relevant reported teleconnections is the ENSO (Nicholson and Kim, 1997; Korecha and Barnston, 2007), whereby most authors found dryer conditions in an El Niño year (Sileshi and Zanke, 2004; Abtew et al., 2009; Segele et al., 2009a, 2009b). Indeed, famines and droughts reported in historical sources (Spinage, 2012) show a small frequency peak at 5–10 years variability (Figure 2). The impact of El Niño is also found to the south of the Horn of Africa. For example, Wolff et al. (2011) found that equatorial East Africa has more rain and flooding during El Niño and droughts during La Niña years. Also, Nicholson and Selato (2000) obtained similar results (reduced rainfall during La Niña) for southern Africa.

Studying climate variability in North Ethiopia is important because of the strong interactions between climatic variability, land degradation and hydrogeomorphic change taking place in the region (Frankl et al., 2011, 2013a). North Ethiopia is often perceived as a region plagued by drought, land degradation and famine. The linkages between climate, land degradation, surface water processes and crop production are quite complex (Rosell, 2011; Frankl et al., 2013b). Ethiopia is sometimes named ‘the water tower of Africa’, as its surface run‐off waters account for about 85% of the Nile flow (Yacob and Imeru, 2005; Gebreyohannes et al., 2013). Notably, land degradation processes, such as severe erosion and water run‐off, are very active in the North Ethiopian Highlands (Nyssen et al., 2004), while often a post‐18th century drying trend (Camberlin, 1994; Yilma and Demarée, 1995) is named as a driving force of such gradual land surface degradation. As a result of agricultural stagnation, Williams and Funk (2011) estimated that because of low food production, more than 15 million people were food insecure in Kenya, Ethiopia and Somalia in 2009. Food insecurity is however a long‐term problem, as famines are reported in Ethiopia since the 12th century (Pankhurst, 1985; Spinage, 2012; summarized in Table 1).

By using rainfall data, the objectives of this study are therefore: (1) to apply an EOT‐based method to detect teleconnections and (2) to understand drought occurrence in North Ethiopia from these teleconnections.

2 Materials and methods

2.3 EOT computations with pressure data

In order to identify large‐scale patterns of pressure anomalies, EOTs were computed with the mean sea level pressure data, using IDRISI software, in a normal setup instead of space–time reversal. EOT analysis provides solutions that are orthogonal in one direction, either space or time, and can be compared to principal components rotated obliquely (van den Dool et al., 2000). Because EOTs are constrained to be orthogonal in just one direction rather than in both space and time directions, the computation is more efficient than is the case with EOFs (Smith, 2004). Contrary to EOF methods, EOT allows some freedom to target the analysis towards specific geographic locations. Moreover, there is no need to determine the degree of EOF truncation before rotation (Simonti and Eastman, 2010). Hence, EOTs were calculated on JJAS mean sea level pressure data, and the EOTs that were significantly correlated with JJAS rainfall in Mekelle were used for analysis. Significance of all correlations was tested with a two‐sided t‐test as described in Snedecor and Cochran (1989):

(1)

Under the null hypothesis there is no correlation between the two variables, n is the number of observations, r the Pearson product moment correlation coefficient and n−2 the degrees of freedom.

Let f(s_l, t_k) be the spatiotemporal pressure data set at locations s_l and at time t_k, and f_μ(s_l, t_k) the mean and trend component; then the EOTs are obtained by computing the pressure anomaly f'(s_l, t_k):

(2)

or:

(3)

with:

(4)

and

(5)

Here, e_p(s_l) is a regression coefficient, corr is the correlation function, σ is the standard deviation, sb_p is the point in space with the most spatial variance at a given time, and c_p(t_k) is the pth temporal eigenfunction (Magar et al., 2012). As an a posteriori validation of the results, existing indices of atmospheric teleconnections were taken into account. This included the JJAS Multivariate ENSO Index (MEI; Rasmusson and Carpenter, 1982; Wolter, 1987), while a JJAS Southwest Monsoon (SWM) Index was used as standardized difference between standardized wind modulus at 925 hPa and standardized zonal wind at 200 hPa, within the West African monsoon domain (5–17.5°N, 20°W–40°E) (Li and Zeng, 2005). A JJAS Western Tropical Indian Ocean Sea Surface Temperature Index was taken in the region (50–70°E; 10°S–10°N). Hence, kiremt rainfall could be linked with teleconnection variables in a multivariate regression.

3 Results

3.1 Identification of the EOTs

From the first 10 EOTs performed on JJAS mean sea level pressure data, three EOTs were significantly correlated with JJAS rainfall at Mekelle‐Quiha Airport: EOT‐1, EOT‐2 and EOT‐5. EOT‐1 represents an Atlantic Ocean signal, EOT‐2 represents a Pacific Ocean signal and EOT‐5 an Indian Ocean signal. These EOTs were also well‐correlated with JJAS rainfall averaged in an area of 125 km around Mekelle (correlation > 0.30), and its spatial patterns (pixels with the most spatial variance) indicate mean sea level pressure focuses on the Gulf of Guinea, the Southern Pacific and the Western Indian Ocean, respectively (Figure 3). Other EOTs (e.g. EOT‐3 or EOT‐4) were not significantly correlated with kiremt rainfall in the study area and were, following Simonti and Eastman (2010), not withheld for further analysis.

EOT‐2 is strongly negatively correlated with the MEI (correlation = −0.67; p < 0.001) and positively with the Southern Oscillation Index (correlation = +0.66; p < 0.001), and is therefore closely related to a La Niña signal. As the EOT‐5 would relate to a (negative) Indian Ocean Dipole (IOD) signal, the variable could be compared to the Dipole Mode Index (DMI). However, while EOT‐5 is hardly correlated with the JJAS MEI (correlation = 0.09) and weakly with the JJAS DMI (correlation = −0.35), it is better correlated with the residuals of the regression JJAS MEI–DMI (correlation = −0.43; p = 0.003). As EOT‐5 does not linearly correlate with MEI, it hence reflects only non‐ENSO modes of variability. The Atlantic EOT is weakly correlated with the JJAS Atlantic Meridional Mode (correlation = 0.22), but better with the JJAS mean sea level pressure over Saint Helena (correlation = 0.38; p = 0.008). EOT‐1 hence should represent some ‘Guinea Anticyclone’ signal, controlling the SWM (Segele et al., 2009b).

3.2 Identification of the interannual rainfall variability

The kiremt time series derived from both NMA as PREC/L do not show any gradual trends, although it is clear that kiremt precipitation underwent some serious dips during the 1980s (Figure 4). During the years 1983–1985, North Ethiopia was struck by severe droughts – these are the years of the ‘Great Famine’. For mean JJAS precipitation, linear and Theil‐Sen trends were tentatively calculated as ‘declining’ (Figure 4), but were not significant for the study area (p > 0.1). Given this quite stable rainfall regime, the rainfall gauge data imply that there is no gradual drying trend over the second half of the 20th century. Moreover, the data confirm other studies based on Ethiopian meteorological stations contesting a presumed gradual 20th century drying trend (e.g. Conway, 2000; Cheung et al., 2008).

3.3 Linking EOTs with interannual rainfall variability

3.3.1 Single kiremt EOT‐logit models

First, logistical regression was performed in order to investigate the impact of the identified teleconnections (Section 3.1) on the probability of having a wet (value of 1) or dry (value of 0) kiremt season. All links between the EOTs and the North Ethiopian precipitation data were significant (Table 2). Hence, such logistic teleconnection–drought models are simple and significant models, yielding for example the following equation for the second teleconnection:

(6)

Table 2. Significance and regression coefficients of all factors in the EOT‐1,2,5 logit models.

	B	Wald	Significance	Exp(B)
EOT‐1	−0.016	7.337	0.007	0.984
Constant	1606.385	7.334	0.007
EOT‐2	0.010	4.738	0.029	5.010
Constant	−0.269	0.695	0.404	0.764
EOT‐5	−0.019	8.001	0.005	0.981
Constant	−0.516	2.058	0.151	0.597

As illustrated by Equation 6, the more negative the EOT‐2 signal (occurrence of a strong El Niño), the higher is the risk of a dry year in North Ethiopia. Further, the logit models showed significantly negative relations with EOT‐1 and EOT‐5.

3.3.2 A multivariate kiremt model

In order to formalize the combined interaction of the teleconnections, JJAS rainfall measured in Mekelle‐Quiha Airport was linked to the teleconnections through a multivariate regression model. As EOT‐1 could not be significantly linked in any linear model but was still significantly correlated with JJAS rainfall, it was replaced by the SWM Index, which gave significant results. Finally, a model with four variables gave significant results (Table 3), which could explain about two third of North Ethiopian rainfall variability. The variables included are the West Tropical Indian Ocean SST Index (WTIO), EOT‐5, the SWM Index, the MEI or EOT‐2. These interactions with the study area are shown schematically in Figure 5. The WTIO is well‐correlated with the MEI (0.48) and with the DMI (0.40). The regression replicates the rainfall measurements rather well (Figure 6(a) and (b)), in particular the droughts during the 1980s. A similar procedure based on multivariate logit regression (Table 3) modelled 83.7% of the drought risks correctly (Table 4).

Table 3. Significance and regression coefficients of all factors in the multivariate models.

Model with EOTs (R2 = 0.62)			Model with indices (R2 = 0.64)			Logit model with indices (Nagelkerke R2 = 0.63)
	B	Significance		B	Significance		B	Significance
Constant	419.16	0.000	Constant	407.59	0.000	Constant	−2.97	0.009
WTIO	313.96	0.000	WTIO	362.63	0.000	WTIO	9.59	0.008
SWM	50.24	0.000	SWM	43.844	0.000	SWM	0.95	0.038
EOT 5	−0.83	0.005	EOT 5	−0.85	0.003	EOT5	−0.025	0.020
EOT 2	0.45	0.018	MEI	−52.89	0.004	MEI	−1.50	0.025

Table 4. Classification matrix of the multivariate logit model.

		Modelled
		0	1	% Correct
Observed	0	23	3	88.5
	1	4	13	76.5
	%			83.7

4 Discussion

The oscillation‐controlled rainfall pattern observed in this study can be partly explained using the conceptual regional climate model of Segele et al. (2009a, 2009b). Through wavelet, regression, correlation and composite analysis, they concluded that enhanced kiremt rainfall in North Ethiopia is dependent on La Niña conditions in particular as Ethiopian rainfall is associated with monsoon low deepening over the Arabian Peninsula. During La Niñas, pressure lowering in the Arabian Monsoon Trough (Figure 7(a)) coincides with pressure strengthening in the Mascarene High, which enhances the Somali Low‐Level Jet and the Tropical Easterly Jet (at 200 hPa). This combination brings more moisture to Ethiopia (Segele et al., 2009b) (Figure 7(b)). Similarly, way to the Northeast of the Horn of Africa, the Asian Monsoons are weakened during El Niño events (Ju and Slingo, 1995). In India, the majority of the warm ENSO (El Niño) episodes are accompanied by dry summers, coincident with low‐level anticyclone anomalies (Lau and Nath, 2000).

Our findings are hence in good agreement with the results of several previous studies (Sileshi and Zanke, 2004; Abtew et al., 2009; Segele et al., 2009a, 2009b), who find that North Ethiopian rainfall is suppressed during El Niños and enhanced during La Niñas. The ability of EOT analysis to identify ENSO conditions is further supported by performing a logistical regression (p = 0.026) with the MEI, giving similar results as the EOT‐2 logit model:

(7)

Contrary to North Ethiopia, equatorial East Africa is wetter during El Niño conditions, because the Intertropical Convergence Zone is intensified by a warmer west equatorial Indian Ocean sea surface temperature (Wolff et al., 2011). There, rainfall occurs from mid‐March to the end of May (the so‐called Long Rains), and from mid‐October to mid‐December (the Short Rains).

Further, Saji et al. (1999) report an impact of the IOD on East African rainfall. During positive IOD events, equatorial winds switch direction and bring warm waters towards the West (Indian Ocean), which enhance a more powerful monsoon (Woods Hole Oceanographic Institution, 2013). However, this relationship was never formalized for the Northern Horn of Africa, and North Ethiopia in particular. In addition, Segele et al. (2009a, 2009b) highlight the importance of the SWM for bringing moisture from the Atlantic Ocean towards the Ethiopian Highlands. As the Congo Air Boundary is situated above Central Ethiopia during July (Bergner and Trauth, 2004), the strength of the Guinea Anticyclone is instrumental in having intensive moisture‐bearing SWM (Segele et al., 2009b). Viste and Sorteberg (2013) confirm the importance of moisture transport from the Gulf of Guinea to Ethiopia. It would be interesting for future research to target specifically the Indian and Atlantic regions and hence ‘refine’ our broad diagnosis.

In addition to the physical mechanisms that regulate the regional rainfall pattern over North Ethiopia, it is clear that oscillation‐induced droughts can have several detrimental effects, as some authors attribute spikes of famine and land degradation to El Niño events. For example, Abtew et al. (2009) conclude that the great Ethiopian famine of 1888–1892 corresponds to a very strong El Nino year (1888). Related to this, Frankl et al. (2011) found that during the late 19th century gullies in Tigray were probably reactivated because of such increased aridity. Equally, the low amplitude interaction between our three teleconnections (EOT‐2, EOT‐5 and SWM) during the years 1982–1983 (Figure 6) is coinciding with the years of Great Famine and severe land degradation in the Ethiopian Highlands. In line with hydrogeomorphological studies, this article formalizes the contribution of oceanic teleconnections to the probability of having a dry year, with consequences for land degradation processes in North Ethiopia. As shown in Figure 6(b), the 1980s droughts can be well‐explained in terms of the three identified teleconnections.

Identifying interactions with quasi‐global teleconnections is important for predictions of future climate change. For East Africa, Williams and Funk (2011) conclude that a warm pool over the Indian Ocean, and the associated movement of diabatic heated air by Walker north‐eastern trade winds towards eastern Africa, will suppress belg precipitation. However, concerning kiremt rains, most climate models predict a wetter East African climate under conditions of global warming (Williams and Funk, 2011), due to a strengthening of the Hadley Circulation. In addition, based on Table 3, an enhanced SWM and increasing Indian Ocean temperatures can further strengthen North Ethiopian rainfall. However, this study illustrated the disruptive character of the El Niño Oscillation that could counter such effects. Hence, future water‐soil‐crop interactions may depend especially on the impacts of global warming on the ENSO, an effect that will need further consideration in Horn of Africa climate projections. At the same time, the temporal pattern of EOT‐5 reveals a significantly positive trend. Taking into account the exceptional heating of the Indian Ocean over the second half of the 20th century (Williams and Funk, 2011), this might be a signal of positive IOD intensification.

5 Conclusions

Focusing on North Ethiopia, this study tried to (1) apply an EOT‐based method to detect teleconnections and (2) understand drought occurrence in North Ethiopia from these teleconnections. Theil‐Sen trend analysis on rain gauge data did not show a declining precipitation trend over the 20th century. However, still interannual variability was found in the data sets. Regression confirmed the impact of La Niña/El Niño conditions, as the second EOT (based on assimilated mean sea level pressures) had a significant impact on kiremt rainfall. Further, the IOD and the SWM are influencing kiremt rainfall variability in North Ethiopia. The interaction between these three teleconnections was found to be a major contributor to the 1983–1985 droughts. As the mutual impacts of the Indian Ocean Warming – IOD on East African rainfall are still less investigated than the impact of ENSO, we argue that East African meteorological services must take care to incorporate IOD interactions into their prediction models.