Abstract

In much of Ethiopia, similar to the Sahelian countries to its west, rainfall from June to September contributes the majority of the annual total, and is crucial to Ethiopia’s water resource and agriculture operations. Drought-related disasters could be mitigated by warnings if skillful summer rainfall predictions were possible with sufficient lead time. This study examines the predictive potential for June–September rainfall in Ethiopia using mainly statistical approaches. The skill of a dynamical approach to predicting the El Niño–Southern Oscillation (ENSO), which impacts Ethiopian rainfall, is assessed. The study attempts to identify global and more regional processes affecting the large-scale summer climate patterns that govern rainfall anomalies. Multivariate statistical techniques are applied to diagnose and predict seasonal rainfall patterns using historical monthly mean global sea surface temperatures and other physically relevant predictor data. Monthly rainfall data come from a newly assembled dense network of stations from the National Meteorological Agency of Ethiopia. Results show that Ethiopia’s June–September rainy season is governed primarily by ENSO, and secondarily reinforced by more local climate indicators near Africa and the Atlantic and Indian Oceans. Rainfall anomaly patterns can be predicted with some skill within a short lead time of the summer season, based on emerging ENSO developments. The ENSO predictability barrier in the Northern Hemisphere spring poses a major challenge to providing seasonal rainfall forecasts two or more months in advance. Prospects for future breakthroughs in ENSO prediction are thus critical to future improvements to Ethiopia’s summer rainfall prediction.

1. Introduction and background

Ethiopia, located within 3.30°–15°N, 33°–48°E, has three climatological rainy seasons: June–September (called Kiremt), October–January (Bega), and February–May (Belg; Shanko and Camberlin 1998; Sileshi and Demarée 1995; Tsegay 1998, 2001; Gissila et al. 2004). Kiremt rains during June–September (JJAS) account for 50%–80% of annual rainfall totals over the regions having high agricultural productivity and major water reservoirs. Thus, the most severe droughts are usually related to a failure of the JJAS rainfall to meet Ethiopia’s agricultural and water resources needs. This study is devoted to Ethiopia’s JJAS rainfall climatology, interannual variability, and predictability.

Tropical rainfall varies from daily, interannual, to interdecadal and longer time scales. Following breakthroughs in weather forecasting in the 1950s and 1960s (see Cane 2000), as environmental monitoring capabilities improved, physical modeling of the interannual variability of sea surface temperature (SST) over the eastern tropical Pacific Ocean revealed predictability of the El Niño–Southern Oscillation (ENSO; Cane et al. 1986; Cane and Zebiak 1987; Zebiak and Cane 1987). ENSO predictability then led to potential predictability of seasonal climate over many tropical and some extratropical regions. Studies have indicated that northern summer rainfall in the Sahel responds partly to ENSO fluctuations (Nicholson and Kim 1997; Nicholson and Selato 2000; Hastenrath 1995; Rowell 2001, among many others). On decadal scales, research has provided evidence and possible explanation for Sahelian drought throughout most of the last quarter of the 1900s (e.g., Hulme 2001). Techniques used at the Met Office, among other global prediction centers, attempt to capture both interannual and interdecadal components of SST forcing. Research devoted to the twentieth-century Sahel drought focused heavily on the impact of regional and global SST anomalies on interdecadal time scales (Folland et al. 1991; Rowell et al. 1995; Ward 1998; Giannini et al. 2003; Janicot et al. 1996; Janicot et al. 2001; Zeng 2003; Paeth and Friederichs 2004). Some studies have addressed the additional influence of land surface forcing (Zeng et al. 1999; Wang et al. 2004).

Sahel droughts have also been studied statistically relative to more regional oceanic and atmospheric factors. Raicich et al. (2003) demonstrated a connection between Indian monsoon and Sahel rainfall regimes and sea level pressure in the Mediterranean area, and Rowell (2003) showed the influence of Mediterranean SSTs on seasonal Sahel rainfall. Osman and Shamseldin (2002) showed that the driest years in central and southern Sudan occur during the warm phase of ENSO and Indian Ocean SST, and proposed empirical rainfall prediction models. Lamb (1977) suggested an extension of Sahel drought toward Ethiopia on the basis of synoptic circulations. In addition, Giannini et al. (2003) attributed the Sahel’s recent drying trend to warmer-than-average low-latitude waters around Africa, which, by forcing deep convection over ocean, decrease monsoon-related continental convergence and rainfall from Senegal to Ethiopia. Such studies that include Ethiopia could be confirmed using gauge rainfalls from a newly assembled dense station network—data that could be included into the Sahel rainfall indices.

Seasonal rainfall patterns over tropical Africa, like the Indian subcontinent, are modulated partly by monsoonal flows (Bhatt 1989; Camberlin 1997). Ethiopia’s rainfall climatology is determined mainly by seasonal changes in large-scale circulation, part of which involves the seasonal north–south movement of the intertropical convergence zone (ITCZ); this resembles what is generally thought to occur in the traditional Sahel region from Sudan to Senegal (Nicholson 1989). The complex orography across Ethiopia shapes the JJAS rainfall patterns spatially and temporally within the season. Year-to-year variability of Ethiopia’s JJAS rainfall patterns has been described in terms of onset, cessation, dry spell occurrences, and growing season duration (Segele and Lamb 2005). Kiremt rainfall advances gradually northward across the western half of the country from March to mid-June, progressing more rapidly across the eastern half from mid-June to mid-July. (From March to May, this rainfall is actually considered part of the Belg rainfall regime.) The mean southwestward retreat of rainfall occurs from early September to November.

The mountain ranges are oriented southwest–northeast, with the Rift Valley bisecting Ethiopia. During JJAS, there are southwest monsoon low-level winds over the Arabian Sea, strong cross-equatorial flow along eastern flank of Africa, and southeasterly trade winds south of the equator (Gissila et al. 2004). With these low-level flows, summer storm development is facilitated by the upper-level tropical easterly jet (TEJ), serving also as westward-steering currents.

JJAS rainfall in the region around Ethiopia is controlled by several climatological features in the lower and upper troposphere (e.g., Hastenrath 1991). These include the following: 1) seasonal northward advance of the ITCZ, persisting over Ethiopia; 2) formation of heat lows over the Sahara and Arabian landmasses; 3) establishment of subtropical high pressure over the Azores, St. Helena, and Mascarene; 4) southerly/southwesterly cross-equatorial moisture flow from the southern Indian Ocean, central tropical Africa, and the equatorial Atlantic; 5) upper-level TEJ flowing over Ethiopia; and 6) low-level jet (Somali jet). Synoptic systems arising from these seasonal circulations have been discussed [Kassahun 1987; Tadesse 1994; (National Meteorological Services Agency) NMSA 1996; Segele and Lamb 2005]. This study will not focus on these local features, per se.

In JJAS, convective activity typically develops over the Ethiopian highlands, while southern and southeastern Ethiopia receive little rain. Using a multidecadal history of rainfall over a dense station network (described below), the spatial distribution of mean total JJAS rainfall (Fig. 1, top) shows the greatest rainfall over the highlands of western/west-central Ethiopia, the northeast and southeast lowlands being relatively dry. Southeastern Ethiopia, closer to East Africa, has rainy seasons during March–May and October–November. The distribution of mean number of days having measurable (≥0.1 mm) rainfall (not shown), follows a similar pattern,1 with western and central Ethiopia receiving measurable rainfall 70%–90% of the days. Farther to the northeast rainfall occurs for only 10–30 days, despite JJAS being the main rainy season. Southern/southeastern Ethiopia receive rains only for a few days in September with the southward retreat of the ITCZ. The bottom of Fig. 1 shows the percentage contribution of JJAS rainfall to the annual total rainfall. Figures 2 and 3 show, respectively, the locations of selected stations, and the seasonal march of mean monthly rainfall at four stations with varying longitude within the central (8°–10°N) latitude band. This study focuses on the central, western, and northern parts of the country that have their main rainy season in JJAS (although stations in the southern/southeastern portion of this focus area may have a mildly bimodal seasonal distribution).

Fig. 1.

(top) Total JJAS rainfall climatology (mm) over Ethiopia, 1971–2000. (bottom)Percentage of 1971–2000 mean total annual rainfall contributed by JJAS rainfall.

Fig. 1.

(top) Total JJAS rainfall climatology (mm) over Ethiopia, 1971–2000. (bottom)Percentage of 1971–2000 mean total annual rainfall contributed by JJAS rainfall.

Fig. 2.

Locations of selected climatological stations.

Fig. 2.

Locations of selected climatological stations.

Fig. 3.

Time series of long-term mean monthly rainfall (mm), 1970–2000, for selected stations within the central (8°–10°N) latitude band. Maximum rainfall occurs in JJAS season, but decreases eastward with bimodal seasonal rainfall patterns more likely over eastern sectors.

Fig. 3.

Time series of long-term mean monthly rainfall (mm), 1970–2000, for selected stations within the central (8°–10°N) latitude band. Maximum rainfall occurs in JJAS season, but decreases eastward with bimodal seasonal rainfall patterns more likely over eastern sectors.

Ethiopian rainfall in JJAS differs vastly from year to year in timing and total amount. The phase of ENSO has been identified as impacting summer rainfall (Nicholls 1993; Tsegay 1998, 2001; Gissila et al. 2004; Segele and Lamb 2005; Sileshi and Demarée 1995; Bekele 1997), with the same direction of impact as that of the Sahel. The large-scale atmospheric dynamics relevant to Ethiopia, however, differ in some ways from those relevant to regions farther west in the Sahel (Bhatt 1989; Cook 1997). Recently, Gissila et al. (2004) developed an empirical forecast model for Ethiopian summer rainfall using regression with Indian and Pacific SSTs in March, April, and May as potential predictors. Our study aims to quantify the statistical relations between ENSO, and other oceanic and atmospheric phenomena, and JJAS rainfall, a practical objective being to develop models that skillfully anticipate rainfall anomalies prior to rainy season onset, allowing for societal mitigation measures.

We utilize more stations than were available for the above studies, and apply several techniques to quantify rainfall behavior with respect to ENSO and other governing large-scale climate patterns, both averaged over all of Ethiopia and distributed geographically within the country. We assess prospects for implementing statistical techniques for reliable, sustainable climate forecasts, including a forecast for the state of the ENSO by a dynamical model as a potential rainfall predictor.

Our datasets and methodologies are described in section 2. ENSO cycles, their implications for predictability of JJAS rainy season, and the predictability of ENSO itself during JJAS, are presented in section 3. Results of diagnostic and predictive rainfall modeling are provided in section 4. Discussion and conclusions are given in section 5.

2. Data and methodology

The rainfall data used for many analyses in this study, obtained from the Ethiopian NMSA, are monthly totals for June–September. Two hundred meteorological stations (Fig. 4a) have periods of records varying from 15 to >50 yr. Nearly half of the full array of stations has records of 30 yr or more since 1961 (Fig. 5). Data from the 1960s are omitted here, due to widespread gaps. A total of 78 stations (Fig. 4b) are used for many of our analyses, usually covering the period 1970–2004.2 However, only 55 stations, denoted by large circles in Fig. 4b, are used for our all-Ethiopian JJAS rainfall analyses, where stations located in the south and southeast lowlands that are climatologically dry during JJAS are excluded. From 1970 onward, the proportion of missing data is low, with a small number of stations having at most 10% missing data (Fig. 6). Missing months were estimated by interpolation from the relative anomalies of stations within a threshold distance away (typically including 1–4 stations). To assess the maximum sensitivity to including these estimated rainfalls, we compared the time series of the standardized all-Ethiopian average JJAS rainfalls resulting from the “cleaned” 55 stations to that using only the 36 stations having full original records. The two versions of the all-Ethiopian rainfall data correlate 0.91, and the largest absolute differences in standardized rainfalls are near 0.7. This could have a visible, although not major, impact on the results. Omitting 35% of the stations needing any treatment is thought to represent an upper limit of the effect of including stations requiring filling of missing data.

Fig. 4.

(left) Locations of the 200 climatological stations over Ethiopia. (right) Same as in (left) but for the 78 stations used in manyof the analyses of JJAS rainfall in Ethiopia, 55 of which (larger filled circles) are used for the analyses of all-Ethiopian rainfall.

Fig. 4.

(left) Locations of the 200 climatological stations over Ethiopia. (right) Same as in (left) but for the 78 stations used in manyof the analyses of JJAS rainfall in Ethiopia, 55 of which (larger filled circles) are used for the analyses of all-Ethiopian rainfall.

Fig. 5.

The number of meteorological stations having nonmissing rainfall data during JJAS as a function of year, used for analyses of JJAS rainfall anomalies during 1961–2004.

Fig. 5.

The number of meteorological stations having nonmissing rainfall data during JJAS as a function of year, used for analyses of JJAS rainfall anomalies during 1961–2004.

Fig. 6.

Number of stations having missing monthly data during JJAS (1970–2004), for the set of 78 stations (upper curve) and the set of 55 stations.

Fig. 6.

Number of stations having missing monthly data during JJAS (1970–2004), for the set of 78 stations (upper curve) and the set of 55 stations.

We use global SST from the National Oceanic and Atmospheric Administration/National Climatic Data Center (NOAA/NCDC) Extended Reconstructed Sea Surface Temperature version 2 (ERSSTv2) historical dataset (Smith and Reynolds 2004), with 2° × 2° resolution for 1970–2004. From these SSTs indices are derived, including the Niño-3.4 ENSO index (SSTs averaged over 5°N–5°S, 120°–170°W). The Niño-3.4 index is used to represent the ENSO condition because of its demonstrated importance for ENSO teleconnections (Trenberth and Hoar 1996; Barnston et al. 1997). Retrospective forecasts for Niño-3.4 SST generated by the Lamont-Doherty Earth Observatory’s dynamical ENSO forecast model (current version LDEO5; Chen et al. 2004) were kindly run by D. Chen for the period 1970–2005.

In examining diagnostic and predictive aspects of summer Ethiopia rainfall, several linear statistical techniques are employed. Standardized anomalies of the 4-month total rainfalls are used for each station for many analyses, using 1971–2000 as the base period. Standardization places data in all locations in a similar frame of reference for assessing year-to-year rainfall anomalies. The drier stations in the Rift Valley, and eastern or southern Ethiopia, tend to have particularly positively skewed JJAS rainfall distributions, and reduced signal-to-noise ratios due to the few governing nondry years. Because these south/southeast lowland stations are in seasonally dry zones, they are omitted from the 55-station subset used for the all-Ethiopian rainfall index (Fig. 4b).

Simple linear correlation, multiple linear regression, and robust regression techniques3 are used to develop predictive models for summer Ethiopian rainfall, revealing teleconnections between SST and rainfall. Strategies to predict the ENSO-related summer Niño-3.4 SST anomaly are used as one approach to forecasting consequent Ethiopian rainfall.

For prediction of JJAS seasonal rainfall, canonical correlation analysis (CCA) is used, as described in previous studies (e.g., Hotelling 1936; Glahn 1968; Barnett and Preisendorfer 1987; Barnston and Smith 1996; Thiaw et al. 1999). CCA is a multivariate regression that relates patterns in predictor fields (e.g., SST) to patterns in a predictand field (e.g., rainfall). Cross validation (Michaelsen 1987) and retroactive designs are used to minimize inflation of the skill estimates. The SST–rainfall relationships are examined both for concurrent data and when SST precedes the summer rainfall, as in actual forecasting.

3. ENSO and the June–September rainy season over Ethiopia

The impact of the ENSO variability on global climate has been well documented (Ropelewski and Halpert 1987; Mason and Goddard 2001; Goddard and Dilley 2005, among many others). The ENSO state modulates the rainy seasons in some regions, particularly in the Tropics (Hastenrath 1995). El Niño is associated with drought and forest fires in parts of Australia, Indonesia, Southeast Asia, and southern Africa (Goldammer 1999; Khandekar et al. 2000; Jury 2002). Chances for flooding are enhanced with El Niño during the short rainy season of October–December in East Africa (Ogallo 1988, 1989; Indeje et al. 2000; Philippon et al. 2002).

During JJAS, suppressed rainfall has been observed to accompany El Niño over much of Ethiopia, often with economic catastrophe. Although the importance of ENSO to Ethiopian rainfall is being accepted and incorporated in the NMSA’s operational policy4 more now than previously, it continues to be somewhat underweighted despite widespread documentation of its importance (NMSA 1996; Camberlin 1997; Bekele 1997; Tsegay 1998; Gissila et al. 2004; Segele and Lamb 2005). As shown in Fig. 7 (for 1970 onward), lower tercile all-Ethiopian JJAS seasonal rainfall occurred in 1965, 1972, 1979, 1982, 1984, 1987, 1990, 1991, 1995, 1997, and 2002. More than half of these summers coincided with El Niño events; none occurred during La Niña. Upper tercile rainfall conditions occurred in 1961, 1964, 1970, 1973, 1974, 1975, 1977, 1978, 1981, 1988, 1994, 1996, 1998, 1999, and 2003; more than half of these matched La Niña events, while only one (1994) occurred with El Niño. Below we will assess this connection more quantitatively, and discuss the potential for issuing useful seasonal rainfall predictions before the onset of summer season rains.

Fig. 7.

Standardized JJAS rainfall anomalies of all-Ethiopian rainfalls for the 1970–2004 period. Years having El Niño, La Niña, and neutral conditions during JJAS, based on the NOAA/CPC ENSO classification, are denoted by the patterns inside the bars.

Fig. 7.

Standardized JJAS rainfall anomalies of all-Ethiopian rainfalls for the 1970–2004 period. Years having El Niño, La Niña, and neutral conditions during JJAS, based on the NOAA/CPC ENSO classification, are denoted by the patterns inside the bars.

The effect of ENSO on rainfall is seen in composite analyses for selected individual stations by month. JJAS monthly rainfalls are averaged for El Niño, La Niña, or near-neutral conditions, using the classification system of the NOAA/Climate Prediction Center (CPC).5 Here, all months of any year are assigned the ENSO phase existing during JJAS of that year, so that impacts of ENSO events occurring during the Belg and Bega seasons are not directly represented. Mean monthly rainfalls seem to be enhanced during La Niña years in regions where JJAS is the major rainy season, due both to greater duration of the rainy season (Segele and Lamb 2005), and increased rainfalls during individual months of the rainy season. Examples of stations from different parts of Ethiopia having a clear ENSO influence are shown in Fig. 8.

Fig. 8.

Seasonal march of mean monthly rainfall amount (mm) composited for years whose JJAS season is classified as El Niño, La Niña, or neutral, for four stations located in the northwest, northeast, central, and southeast portions of Ethiopia, respectively: Bahir Dar (labeled BDR in Fig. 2; northwest), Addis Ababa (AAB; central), Bati (BTI; northeast), and Gode (GDE southeast; influence in both MAM and OND seasons). Based on 1970–2004 data.

Fig. 8.

Seasonal march of mean monthly rainfall amount (mm) composited for years whose JJAS season is classified as El Niño, La Niña, or neutral, for four stations located in the northwest, northeast, central, and southeast portions of Ethiopia, respectively: Bahir Dar (labeled BDR in Fig. 2; northwest), Addis Ababa (AAB; central), Bati (BTI; northeast), and Gode (GDE southeast; influence in both MAM and OND seasons). Based on 1970–2004 data.

Figure 9 illustrates the geographical distribution of the correlation between the SST in the Niño-3.4 index region and Ethiopian JJAS rainfall at the 78 stations, based on 1970–2004, keying SST to individual months prior to summer (Figs. 9a–c) and SST during JAS (Fig. 9d). The association of summer rainfall with ENSO in early presummer months (January–April) is weak, and increases as the time of the ENSO state approaches the beginning of the rainfall season. Statistically significant (≥0.34) negative correlations are found between JJAS rainfall totals and Niño-3.4 SST occurring nearly simultaneously (in JAS) mainly in the northern half of the country but also in the southern highlands and southwest Ethiopia (Fig. 9d). In the climatologically dry southeastern lowlands, associations with ENSO are weak. The moderate negative simultaneous correlations (−0.4 to −0.6 at some locations) imply that rainfall forecasts would have useful skill levels if the summer Niño-3.4 SST could be predicted beforehand. Correlations between JJAS rainfall and Niño-3.4 SSTs of preseason months may be of some use only for May, where some correlations are stronger than −0.4. The lack of a stronger relationship between the May ENSO state and rainfall is not surprising, as the ENSO condition may change in either direction between April and June (Tziperman et al. 1998). For example, high Niño-3.4 SST in May could be due to an El Niño that had matured earlier and would likely dissipate before July, or to a newly emerging El Niño that was absent in February and March. Predicting ENSO is known to be difficult during the northern spring. Later we will discuss an indicator of summer ENSO based on the change of the May SST anomaly from that of a few months earlier.

Fig. 9.

Spatial distribution of correlation between JJAS rainfall for 78 stations in Ethiopia and Niño-3.4 SST in (a) January, (b) April, (c) May, and (d) JAS. Computed for 1970–2004, values of 0.34 or greater in magnitude are statistically significant at the 95% confidence level.

Fig. 9.

Spatial distribution of correlation between JJAS rainfall for 78 stations in Ethiopia and Niño-3.4 SST in (a) January, (b) April, (c) May, and (d) JAS. Computed for 1970–2004, values of 0.34 or greater in magnitude are statistically significant at the 95% confidence level.

A time series of the all-Ethiopian JJAS rainfall average is derived for 1970–2004, based on the above-mentioned 55 stations (Fig. 4b). The time series of the average of the seasonal rainfall totals, restandardized by 1971–2000 rainfall statistics, is shown in Fig. 10. Overall deficient (abundant) rainfall tends to occur during El Niño (La Niña) summers, the four strongest for JJAS being 1972, 1982, 1987, and 1997 (1973, 1975, 1988, and 1999). The correlation with Niño-3.4 SST over 35 yr (1970–2004) is −0.76.6

Fig. 10.

Standardized JJAS rainfall anomalies of (top) all-Ethiopian rainfalls and (bottom) those of Niño-3.4 SSTs for 1970–2004. Standardization based on 1971–2000 statistics. Correlation between the two is −0.76.

Fig. 10.

Standardized JJAS rainfall anomalies of (top) all-Ethiopian rainfalls and (bottom) those of Niño-3.4 SSTs for 1970–2004. Standardization based on 1971–2000 statistics. Correlation between the two is −0.76.

Interestingly, although 1997 marked the strongest El Niño during 1970–2004, the JJAS all-Ethiopian rainfall was only the sixth lowest among the 35 yr. This is explained partly by the modulating roles of other tropical ocean basins in governing Ethiopia’s JJAS rainfall. In 1997 the Indian Ocean’s delayed warming response to the abnormally warm tropical Pacific (e.g., Goddard et al. 2001) appeared earlier than normal during late northern summer/fall, due both to the seasonally early onset of Pacific warming (April 1997) and the magnitude of that warmth. Thus, in September 1997, with a noticeably warmed western Indian Ocean SST, a northward meridional extension of the ITCZ induced widespread rains over Ethiopia. (During most El Niño northern summers, the Pacific, but not yet the Indian Ocean, has warmed.) In other severe summer droughts, such as the ENSO-neutral 1984, anomalously warm SST was present in the eastern equatorial Atlantic, with an attendant southward retreat of ITCZ and similar displacement of the monsoon trough (Lamb 1978; Ward 1998; Segele and Lamb 2005).

Furthermore, the upper-tropospheric TEJ (Camberlin 1997) supports Ethiopian rainfall in JJAS (Segele and Lamb 2005). A strong TEJ, which is consistent with above-average SSTs in the northwestern tropical Pacific (and thus, indirectly, with La Niña conditions), was observed in summer 1996, as opposed to a poor TEJ in the dry year of 1984 despite a cool/neutral ENSO state. A comparable positive association between TEJ strength/latitude and Sahel monsoon rainfall was identified by Hastenrath (2000). TEJ and other atmospheric systems may be partly associated with the ENSO state—again implying the importance of skillfully predicting the JJAS ENSO condition. How well can the summer ENSO condition be predicted upon completion of May, just in time to anticipate summer rainfall?

The strengths of linear relationship between all-Ethiopian JJAS rainfalls and the Niño-3.4 SST index for individual months from January to September, and for JJA and JAS SST, are shown in Table 1 for 1970–2004. Correlations are near −0.75 during the months of the summer rainy season, stronger than correlations presented in Fig. 9d for any individual station, due to the filtering effects of spatial aggregation with respect to the random variability present in single location rainfalls (Gong et al. 2003). Such noise filtering better isolates the ENSO signal. The correlation is moderate (−0.59) for the May Niño-3.4 SST, suggestive of some predictability based solely on the May ENSO state. A categorical version of the JJAS simultaneous relationship between ENSO and rainfall is shown by a contingency table (Table 2). A moderately strong degree of categorical association between the ENSO and rainfall category is clear, with impacts for both El Niño and La Niña. A chi-square test yields >99% statistical significance.

Table 1.

Correlation, based on 1970–2004, between all-Ethiopian JJAS rainfall and Niño-3.4 SST during monthly or three-month periods prior to and concurrent with the rainfall.

Correlation, based on 1970–2004, between all-Ethiopian JJAS rainfall and Niño-3.4 SST during monthly or three-month periods prior to and concurrent with the rainfall.
Correlation, based on 1970–2004, between all-Ethiopian JJAS rainfall and Niño-3.4 SST during monthly or three-month periods prior to and concurrent with the rainfall.
Table 2.

Association between the ENSO state and all-Ethiopian JJAS rainfall, based on the 1970–2004 period. Table entries are observed frequencies, followed in parentheses by their inferred conditional probabilities (×100), given the ENSO category. Rainfall categorization is based on the three categories having cutoffs at ±0.431 (tercile defining) standardized anomalies. ENSO classification is taken from NOAA/CPC. For the 35 yr of JJAS (1970–2004), 9 (8) years are classified as El Niño (La Niña), and 18 yr as neutral. Significance is assessed using a chi-square test.

Association between the ENSO state and all-Ethiopian JJAS rainfall, based on the 1970–2004 period. Table entries are observed frequencies, followed in parentheses by their inferred conditional probabilities (×100), given the ENSO category. Rainfall categorization is based on the three categories having cutoffs at ±0.431 (tercile defining) standardized anomalies. ENSO classification is taken from NOAA/CPC. For the 35 yr of JJAS (1970–2004), 9 (8) years are classified as El Niño (La Niña), and 18 yr as neutral. Significance is assessed using a chi-square test.
Association between the ENSO state and all-Ethiopian JJAS rainfall, based on the 1970–2004 period. Table entries are observed frequencies, followed in parentheses by their inferred conditional probabilities (×100), given the ENSO category. Rainfall categorization is based on the three categories having cutoffs at ±0.431 (tercile defining) standardized anomalies. ENSO classification is taken from NOAA/CPC. For the 35 yr of JJAS (1970–2004), 9 (8) years are classified as El Niño (La Niña), and 18 yr as neutral. Significance is assessed using a chi-square test.

If the ENSO state for JJAS could be predicted well in advance, much could be said about the general character of Ethiopia’s main rainy season. Since Zebiak and Cane (1987) first established a successful simplified but fully physical coupled ocean–atmosphere model for forecasting ENSO, copious research has been conducted to improve understanding of ENSO and to predict it at several seasons lead (Latif et al. 1998). ENSO forecasts whose lead time traverses the April–June period are known to have lower skill than forecasts whose lead time does not include that period. The seasonal variation of the persistence of tropical Pacific SST anomalies roughly parallels that of the skill (Wright et al. 1988)—autocorrelation of tropical Pacific SST anomalies at 2–4 months is lowest in boreal spring and highest in fall (Latif et al. 1998). Lag correlations between northern summer tropical Pacific SST from the preceding months represent a lower limit of ENSO-related predictability, and we expect autocorrelations of the ENSO state prior to summer with that of summer to weaken as the lead time increases, paralleling the weakening relationship between all-Ethiopian JJAS rainfall and tropical Pacific SST as the time of the SST retreats from JAS to May, April, etc. Our result (Table 3) confirms declining autocorrelations for presummer months, and thus poor relationships using January, February, or March. May is better autocorrelated (0.6–0.7) with the following individual months and the three-month mean summer SSTs. Thus, in anticipating the summer ENSO condition based on earlier ENSO conditions, May SST anomalies have moderate utility, April’s anomalies are weakly helpful, and earlier anomalies are virtually useless.

Table 3.

Autocorrelation (×100) for the Niño-3.4 SST index for 1970–2004. Correlations of 0.60 or higher for periods during the JJAS rainy season from months prior to JJAS are shown in bold.

Autocorrelation (×100) for the Niño-3.4 SST index for 1970–2004. Correlations of 0.60 or higher for periods during the JJAS rainy season from months prior to JJAS are shown in bold.
Autocorrelation (×100) for the Niño-3.4 SST index for 1970–2004. Correlations of 0.60 or higher for periods during the JJAS rainy season from months prior to JJAS are shown in bold.

Lag correlation for the Southern Oscillation index (SOI), an atmospheric component of ENSO, produces similar results. The SOI may be used in tandem with SST for a more balanced and complete ENSO representation. However, monthly SOI, derived from the sea level pressures of two stations, is “noisier” than the monthly SST index.

Methods to predict the summer ENSO state beyond simple SST autocorrelation could involve dynamical models, or statistical models using physically based predictors. We explore both approaches. The intermediate coupled ocean–atmosphere ENSO prediction model originally developed at Lamont-Doherty Earth Observatory (Zebiak and Cane 1987), with current version LDEO5 (Chen et al. 2004), uses sea level, winds, and SST to initialize the predictions. Hindcasts of Niño-3.4 SSTs for JAS were run using reconstructed initialization data from a start time of June 1 (i.e., data through May) for each summer from 1970 to 2004. These hindcasts achieve a correlation of 0.9 against observations.

To determine whether a purely statistical model based on observed data through May can attain similar hindcast skill, we develop linear models to predict the JAS ENSO state. Historical records of Niño-3.4 SSTs and SOI over 1951–90 are used to develop the model. Model selection is conducted considering individual or collective SST and SOI values of January to May as predictors, to predict JAS Niño-3.4 SST. Stepwise multiple linear regression is applied to select predictors, stopping when additional predictors no longer significantly enhance predictive skill. The resulting model used three predictors: 1) May SOI, 2) May Niño-3.4 SST, and 3) May Niño-3.4 SST anomaly minus the February–March average Niño-3.4 SST (MFMSST). This model results in a highly significant multiple R2 of 0.68 (adjusted R2 of 0.66). The model equation, with standardized variables, is

 
formula

The time series of hindcasts from LDEO5 and the multiple regression model are shown in Fig. 11. The regression forecasts are shown both within the training period (1951–90) and for an independent (1991–2004) verification period. The statistical model performance is slightly lower than that of LDEO5, as noted from the prediction differences for 1998, 2003, and 2004. The skills of both tools clearly indicate predictive utility for summer ENSO from the end of May. Even this short lead time would be valuable for early warning of a shift in the odds for JJAS rainfall anomalies. In actual practice, the regression tool can be used when more computer-intensive dynamical or more advanced statistical ENSO forecasts are inaccessible.

Fig. 11.

Time series of the three-month mean (JAS) Niño-3.4 SST anomalies (°C) as observed for the periods 1951–2004, the model simulated using LDEO5, and the multiple linear regression model fitted in the present study. LDEO5-predicted SSTs were available for 1970–2004, whereas the multiple linear regression model is built based on 1951–90 and validated for the remaining period.

Fig. 11.

Time series of the three-month mean (JAS) Niño-3.4 SST anomalies (°C) as observed for the periods 1951–2004, the model simulated using LDEO5, and the multiple linear regression model fitted in the present study. LDEO5-predicted SSTs were available for 1970–2004, whereas the multiple linear regression model is built based on 1951–90 and validated for the remaining period.

4. Statistical rainfall predictions

Now we describe results of prediction schemes to forecast Ethiopian rainfall—first by exploring patterns of correlation with global SST, then using multiple regression and CCA as tools. We consider the simultaneous (summer) relationships as well as those using the states of the predictors prior to onset of the rainfall season.

Because SST anomalies of the global tropical oceans, particularly ENSO, are known to physically induce shifts from the climatologically expected JJAS rainfall probability distribution over Ethiopia, we anticipate certain features in the geographical distribution of correlation between all-Ethiopian JJAS rainfall and SST during May (Fig. 12, top) and August (bottom). August is used to represent the summer SSTs, given that SST anomalies usually change slowly. The distribution of correlation with August SST is seen to be roughly an amplified version of that with May SST. This makes sense, as it is the concurrent SSTs that most directly affect Ethiopian summer rainfall, and the May SST anomaly patterns often resemble those of August, given a moderate three-month autocorrelation. If an early warning procedure were based on this tool, updates in June and July would incorporate further evolution in the SST anomaly pattern after May.

Fig. 12.

Correlation between SST anomalies for (top) May and standardized all-Ethiopian JJAS rainfall anomalies observed from 1970–2004. Contour interval is 0.1. (bottom) Same as (top), but for August SST. Correlation magnitudes of 0.34 and more are statistically significant at 95% confidence level.

Fig. 12.

Correlation between SST anomalies for (top) May and standardized all-Ethiopian JJAS rainfall anomalies observed from 1970–2004. Contour interval is 0.1. (bottom) Same as (top), but for August SST. Correlation magnitudes of 0.34 and more are statistically significant at 95% confidence level.

The most obvious feature in Fig. 12, both for May and August SST, is ENSO related, with the positive (negative) ENSO phase associated with low (high) seasonal rainfall. The correlation patterns over the Atlantic and Indian Oceans do not show strong features. Some positive correlation between rainfall and May SSTs appear in the off-equatorial western tropical Pacific, the southeast Indian Ocean, and weakly in the equatorial Atlantic. Negative values in the central-eastern tropical Pacific are stronger for August than for May SSTs, as are positive correlations near Indonesia and the far eastern Indian Ocean. Weak negative correlations are noted in the subtropical South Atlantic for May SST.

We know that the southwestern Indian Ocean supplies moisture for Ethiopian rainfall during JJAS through a west–east-oscillating Mascarene high pressure center climatologically positioned near the Mascarene Islands during northern summer. The mean summer 1000- and 850-hPa wind flow for the 1970–99 period (Fig. 13) shows Mascarene and St. Helena high pressure, centered near 25°S, 65°E and 25°S, 5°W, respectively. The moist air north of the Mascarene high is forced northward through central equatorial Africa, finally reaching northern Ethiopia. The Congo–moist air boundary is a transient quasi-meridional discontinuity formed by converging winds from the Mascarene and St. Helena highs, and pumps moist air toward Ethiopia. A negative SST anomaly near the Mascarene Islands enhances the Mascarene high, increasing cross-equatorial moisture flow toward Ethiopia. Weak negative SST correlations appear in the Mascarene high region in Fig. 12.

Fig. 13.

Mean summer wind flow and locations of prominent seasonal synoptic systems, based on 1970–99, for (top) 850 and (bottom) 1000 hPa, from NCEP–NCAR reanalysis. (Figure contributed by Z. T. Segele)

Fig. 13.

Mean summer wind flow and locations of prominent seasonal synoptic systems, based on 1970–99, for (top) 850 and (bottom) 1000 hPa, from NCEP–NCAR reanalysis. (Figure contributed by Z. T. Segele)

A weak negative correlation over the equatorial northeast tropical Atlantic, off the West Africa coast, may reflect the SST’s role in modulating the strength and extent of the ITCZ’s northward migration. Warm SST there encourages airflow (and moisture) both from Atlantic and the central Africa rain forest (Congo–moist air boundary) toward the warm pool, depriving Ethiopia of rainfall. The cloudiness off the coast of West Africa often later causes the positive SST anomalies to become negative.

Positive SST correlations appearing during May and August over the western tropical Pacific, related partly to ENSO, may have consequences for Ethiopian summer rainfall through the 100–200-hPa TEJ. The northwestern tropical Pacific and South China Sea are source regions for TEJ, which tends to be stronger during positive SST anomalies in those waters. Various studies (e.g., Segele and Lamb 2005; Kidson 1977; Ward 1998; Grist et al. 2002) do not directly examine an association of TEJ strength with ENSO, or with JJAS Ethiopian rainfall. However, Grist et al. (2002) and Ntale et al. (2003) indicate roles for easterly waves and TEJ, linking them to both ocean forcing and seasonal rainfall for West and eastern equatorial Africa, respectively, for their rainy seasons. Formal quantification of the role of TEJ in Ethiopia’s Kiremt rainfall, and its association with ENSO, remains open.

All-Ethiopian-average JJAS rainfall can be expressed as a linear combination of atmospheric and oceanic predictors whose values are available upon completion of May. We select this short lead time because much of the rainfall predictability comes from the ENSO state expected during summer, and the evolution of this state is difficult to identify earlier than the end of May. A linear regression model is used, candidate predictors being a selected subset of the available SST data including Niño-3.4 SST, SST over part of the southern, tropical and northern Atlantic sectors, the southwest and northwest Indian Ocean, and the SOI. We consider both May values and the changes of the predictors between February–March and May.

The stepwise regression, after passage of diagnostics related to the fitting and the model assumptions, accepts three predictors: 1) the difference of May minus the February–March SSTs over the south Atlantic in the box defined by 30°–40°S, 15°–30°W (MFM_SA); 2) the difference of May minus the February–March Niño-3.4 SST (MFM_Niño-3.4); and 3) May Niño-3.4 SST (May_Niño-3.4). The regression did not select SSTs in the southwest Indian Ocean, despite its above-mentioned possible role and its importance for rainfall in parts of Ethiopia as illustrated for simultaneous Indian Ocean SST–rainfall correlations in Gissila et al. (2004). Standardizing all variables, the model equation is

 
formula

All coefficients are statistically significant at the 95% level, and the overall model “goodness of fit” is significant at 99%. The model explains 59% of the total variance of all-Ethiopian JJAS seasonal rainfall (R = 0.77); the adjusted R2 is 0.53 (R = 0.73). Diagnostic analysis of the model reveals approximate normality of all variables (including rainfall), minimal serial correlations of the residuals, and only mild outlier presence.7

The coefficient estimates remained stable when using cross validation (Michaelsen 1987) and retroactive validation methods (e.g., Barnston et al. 1994)—both widely used in climate prediction (e.g., Thiaw et al. 1999; Mutai et al. 1998, Gissila et al. 2004). In cross validation, a model is developed using all years but excluding each single year, in turn, which is predicted and verified in each case. The retroactive method involves partitioning the time series data into a training period and an independent verification period. We use 1970–96 for training and 1997–2004 for verification.

Results using both hindcasting techniques are shown in Fig. 14. Models fitted and verified using either cross validation or retroactive techniques performed well on most wet years but underestimated the severity of some dry years. The skill of the cross-validation design is superior to that using the retroactive approach; R2 values are 0.41 (R = 0.64) and 0.26 (R = 0.51), respectively. The smaller set of years validated in the retroactive scheme may have rendered its result a less stable estimate of the expected skill. Also, the 1-yr-out style of cross validation can produce still slightly inflated skills, especially when there is high serial correlation in the data (not found in our case). In the cross-validation results, the differences between the model coefficients developed with differing years withheld are small.

Fig. 14.

Predicted and observed all-Ethiopian JJAS standardized rainfall anomalies. The three multiple linear regression predictors, shown in the equation above, include Niño-3.4 SSTs and Atlantic SSTs for the months prior to the beginning of the JJAS rainy season. The retroactive model is fitted to the data series over the training period, then validated for 1997–2004 (thin gray line). The cross-validation model predictions (thicker gray line), and the observations (black line), are shown.

Fig. 14.

Predicted and observed all-Ethiopian JJAS standardized rainfall anomalies. The three multiple linear regression predictors, shown in the equation above, include Niño-3.4 SSTs and Atlantic SSTs for the months prior to the beginning of the JJAS rainy season. The retroactive model is fitted to the data series over the training period, then validated for 1997–2004 (thin gray line). The cross-validation model predictions (thicker gray line), and the observations (black line), are shown.

When fitting the relationship between SSTs and rainfall concurrently in time, the JAS Niño-3.4 SSTs alone explain 58% of the variance of the JJAS all-Ethiopian rainfall (correlation 0.76)—far more than any other candidate predictors, and a one-predictor (simple) regression is sufficient (Fig. 15). Gissila et al. (2004) identified the main regions of SST concurrently associated with Ethiopian rainfall as being the western Indian Ocean, the eastern Indian Ocean, and the eastern tropical Pacific. They assumed that these regions would remain important when used predictively, for individual clustering-determined homogeneous geographical sectors of Ethiopia. Our omission of Indian Ocean SST predictors may be attributed to use of a single all-Ethiopian rainfall index, and, more likely, our focus on the predictive rather than the concurrent SST correlations (top rather than bottom in Fig. 12). Recall that when an ENSO episode develops, typically the Indian Ocean SST anomaly has not yet responded to the growing anomaly in the tropical Pacific during May or June.

Fig. 15.

Model predicted and actual all-Ethiopian JJAS standardized rainfall anomalies. The predictor is JAS Niño-3.4 SST anomalies (JAS_Niño-3.4) that are fitted to JJAS rainfall by simple linear regression. The regression model is statistically significant with confidence levels of 99%. The JAS Niño-3.4 SSTs alone explain 58% of the total variance of JJAS rainfall.

Fig. 15.

Model predicted and actual all-Ethiopian JJAS standardized rainfall anomalies. The predictor is JAS Niño-3.4 SST anomalies (JAS_Niño-3.4) that are fitted to JJAS rainfall by simple linear regression. The regression model is statistically significant with confidence levels of 99%. The JAS Niño-3.4 SSTs alone explain 58% of the total variance of JJAS rainfall.

Skill in predicting all-Ethiopian JJAS rainfall is also assessed using CCA.8 CCA identifies linear relationships between predictor and predictand in a manner similar to multiple linear regression, except that CCA is multivariate on both the predictor and the predictand sides and thus accommodates coupled spatial patterns linking the two fields (or two sets of fields). CCA involves eigenanalysis, in that a matrix of correlations between only cross-dataset (predictor–predictand) elements is processed and then subjected to empirical orthogonal function (EOF) analysis. Here, to reduce noise and the potential for overfitting, the predictor and predictand data individually are pre-orthogonalized using ordinary EOF analysis before applying the CCA, and the CCA then receives the amplitudes of just a few EOFs of the predictor versus a few EOFs of the predictand fields (Barnett and Preisendorfer 1987; Barnston and Smith 1996; Ward 1998; Thiaw et al. 1999). Here we perform CCA starting with 1970–2002 gridded global sea surface temperatures as the predictor field and the set of Ethiopian JJAS standardized individual station rainfall anomalies as the predictand field. Based on skill trials using cross validation, we retain only two modes of predictor and predictand in the pre-orthogonalization, and also two modes in the CCA itself. In two separate rounds of CCA, we first use May and then use JAS SSTs as the predictor, in each case using JJAS rainfall over Ethiopia as the predictand. Time evolution within the predictor SSTs, as might be captured by the change in SST anomalies from February–March to May, is not incorporated. An advantage of CCA is that patterns in the SST field (not just discrete SST index values), are related to patterns in the spatial distribution of Ethiopian rainfall anomaly (not just a single Ethiopia average anomaly). If most of the Ethiopian stations tend to have mutually coherent anomalies, as would be the case if they all respond similarly to ENSO, then the benefit of CCA’s pattern accommodation may not be pronounced on the predictand side.

A CCA based on 1970–2002 data between May SSTs and the set of Ethiopia station JJAS rainfalls—a short-lead forecast—produces results as shown in Fig. 16. The pattern of May SSTs giving rise to skill in predicting the rainfall pattern over Ethiopia is shown by the spatial loading map for the leading CCA mode (Fig. 16, top). This pattern shows a positive ENSO phase (El Niño), associated with mainly negative JJAS rainfall anomalies over Ethiopia as seen in the predictand loading pattern (Fig. 16, middle). The bottom panel shows the temporal scores of the predictor and predictand associated with this mode. The May SST pattern appears to have good skill in capturing the seasonal rainfall performance observed during strong El Niño years (e.g., 1972, 1982, 1987, and 1997) and the strongest La Niña years (1973, 1975, 1988, and 1998). The canonical correlation, describing the strength of the relationship between the predictor SST patterns and predictand rainfall patterns (i.e., the correlation between the predictor and predictand temporal scores), is 0.55.

Fig. 16.

Spatial loadings of the first CCA mode (called EOF1) for the prediction of (middle) Ethiopian JJAS station rainfall, based on (top) May SST. (bottom) The time series of the temporal scores of the predictor SSTs (green line) in predicting JJAS rainfall (red line) for this CCA mode. Based on 1970–2002 data.

Fig. 16.

Spatial loadings of the first CCA mode (called EOF1) for the prediction of (middle) Ethiopian JJAS station rainfall, based on (top) May SST. (bottom) The time series of the temporal scores of the predictor SSTs (green line) in predicting JJAS rainfall (red line) for this CCA mode. Based on 1970–2002 data.

When the JAS SST field is the “predictor” (Fig. 17), the SST predictor loading pattern has a stronger tropical Pacific ENSO pattern, and includes the Indian Ocean to a greater degree than for May SST. This would be expected, given the lagged response of the Indian Ocean with respect to the tropical Pacific Ocean during ENSO episodes (Goddard and Graham 1999; Goddard et al. 2001). While the predictand rainfall loading pattern is very similar to the result using May SST, a stronger correspondence between predictor and predictand temporal scores is evident (Fig. 17, bottom), with a canonical correlation of 0.70. This implies that the May and JAS SST patterns are associated with nearly the same JJAS rainfall response over Ethiopia, but confidence is greater for JAS SSTs. Inclusion of the second CCA mode (not shown) does not improve predictive skill materially, as mode 2 accounts for far less variance than mode 1 for both May and JAS SST predictors. The morphologically similar result seen for concurrent SST–rainfall patterns and the lagged relationship suggests again that the Ethiopian rainfall anomaly pattern is primarily related to the accompanying summer ENSO state, and that successful rainfall prediction depends critically on prediction of that ENSO state from an earlier time. The regression results for predicting ENSO, presented above, demonstrated the need to ascertain the direction of change of ENSO from an earlier month (e.g., February or March) to May, to determine whether the May ENSO condition is in a growth stage or a dissipative stage relative to the previous 1-yr ENSO cycle.

Fig. 17.

Same as in Fig. 16 [spatial loadings of (middle) Ethiopian rainfall of (top) SST, and (bottom) temporal scores for each], except that the SST is for JAS.

Fig. 17.

Same as in Fig. 16 [spatial loadings of (middle) Ethiopian rainfall of (top) SST, and (bottom) temporal scores for each], except that the SST is for JAS.

The temporal scores for the leading CCA predictor mode using either the May or JAS SST field against Ethiopia’s JJAS rainfall station network were each used separately as the single predictor in a simple regression with the all-Ethiopian rainfall as a scalar predictand. The two simple linear regression models generate predicted rainfall time series as shown in Fig. 18. The predictions generated from the 1-mode CCA using JAS SST bear close resemblance with the results for multiple regression using a few scalar predictors (Fig. 15). This indicates that the summer ENSO state is close to being the optimum single predictor for JJAS Ethiopian rainfall, whether the rainfall is described in a single index or as a more detailed anomaly pattern across the stations, and whether ENSO is captured in a single SST index or as a detailed spatial pattern. By contrast, the May SST pattern of the leading CCA mode predicts the all-Ethiopian JJAS rainfall anomalies much less well than the multiple regression whose result was shown above in Fig. 14. This shows that the May ENSO state alone, even as a detailed SST spatial pattern, cannot lead to as skillful a JJAS rainfall prediction as the May ENSO state plus other critical quantities—most importantly, the change of the ENSO state between an earlier month and May. A CCA that uses both the February–March SST and the May SST as temporally “stacked” predictor fields might match or exceed the multiple regression skill; such a CCA design could be explored in future research.

Fig. 18.

Model simulated (lines with circles and triangles) and observed (line with diamonds) all-Ethiopian JJAS standardized rainfall anomalies, based on May and JAS CCA mode 1. The simple linear regression is fitted over 1970–96. The skills, expressed as percentages of total variance explained, of all-Ethiopian JJAS rainfall are 25% (May) and 49% (JAS).

Fig. 18.

Model simulated (lines with circles and triangles) and observed (line with diamonds) all-Ethiopian JJAS standardized rainfall anomalies, based on May and JAS CCA mode 1. The simple linear regression is fitted over 1970–96. The skills, expressed as percentages of total variance explained, of all-Ethiopian JJAS rainfall are 25% (May) and 49% (JAS).

5. Discussion and conclusions

In most of Ethiopia, adequate rainfall during the main rainy season (JJAS) is essential for major societal operations such as hydropower generation, agricultural irrigation, and drinking water. This study examines the potential for predictions of JJAS rainfall with a lead time sufficient for proactive risk management.

A rainfall climatology is derived from a newly assembled dense network of Ethiopian stations for the 1970–2004 period. Previous studies suggested that ENSO-related SST anomalies have a predictable and physically based effect on Ethiopian JJAS rainfall. Here, using the new station rainfalls, we examine the potential to predict JJAS Ethiopian rainfall based on the climate state prior to the onset of the rainy season using statistical techniques, and exploring the skill of statistical methods and one dynamical method to predict the all-important summer ENSO state.

A moderately strong teleconnective relationship between the northern summer ENSO state and concurrent JJAS Ethiopian seasonal rainfall is demonstrated, La Niña (El Niño) associating with enhanced (suppressed) summer rainfall across much of the country. Six out of the nine El Niño years in the 1970–2004 period have been in the dry tercile of the all-Ethiopian JJAS rainfall distribution, while seven out of the eight La Niña years have been in the wet tercile. The ENSO response is strongest over the northern half of the country where the rainfall patterns often depend on the northward advance of the ITCZ during northern summer (Segele and Lamb 2005; Tsegay 2001; Fraedrich et al. 1997). A relationship between the seasonal oscillation of ITCZ and ENSO, and consequences for Ethiopia’s JJAS rainfall, has been suggested (Degefu 1987; Sileshi and Demarée 1995; Segele and Lamb 2005). One linkage scenario involves a weakening and retreating of ITCZ as El Niño episodes begin maturing during late northern summer; the converse would occur in the case of La Niña. In particular, the northward-protruding meridional arm of ITCZ associated with rainfall in central and northern Ethiopia, depending on cross-equatorial northward flow of moist air, may be affected by ENSO.

It is found possible to use the presummer ENSO state, and its direction and rate of evolution, as a simple statistical precursor for the ENSO state during the coming summer season, and consequently the summer seasonal rainfall.

Because of the changeable and uncertain evolution of the ENSO state during northern spring season (the ENSO “predictability barrier”), the strength of association between ENSO and JJAS rainfall decreases sharply as the time of the ENSO state retreats to progressively earlier months—particularly from May backward. The May ENSO state alone provides some indication of the summer Ethiopian rainfall, but the temporal change of May SST from a few months earlier is an essential additional predictor for JJAS ENSO (and hence rainfall), discriminating between growing and decaying ENSO episodes. The LDEO5 intermediate dynamical ENSO prediction model is found to produce skillful ENSO forecasts for the northern summer season using initialized SST data through the end of May. Simple statistical models based on historical Niño-3.4 SST index and SOI in May, and the change from several months earlier, are also shown to produce skillful forecasts of the July–September (JAS) Niño-3.4 SSTs. This simple model could be used in the absence of significant resources and would be further enhanced by merging it with outputs from dynamical ENSO forecasting models such as LDEO5 or others. The northern spring barrier is more than halfway traversed by the end of May and a moderately skillful summer forecast can be made at this short lead time. When and if ENSO can be better predicted through this difficult time of year, longer lead forecasts could be made for Ethiopian summer rainfall.

Rainfall teleconnections to SST regions other than the tropical Pacific are considerably weaker and of smaller spatial scale, and include the Indian and Atlantic Oceans both during and preceding summer.

Multiple linear regression and CCA models are developed to predict JJAS rainfall directly, without predicting the summer ENSO state explicitly. Multiple regression is applied to all-Ethiopian JJAS rainfall, using SST indices and SOI as predictors. The stepwise design selected May Niño-3.4 SSTs, its recent time derivative, and the recent time derivative of SSTs in the subtropical South Atlantic, explaining 59% of the interannual all-Ethiopian JJAS rainfall variance. Pertinent to the key role of the ENSO state to occur during the summer season, the JAS Niño-3.4 SSTs can be used as an alternative “predictor” (after being predicted earlier) that alone would have a better predictive skill score than the above three precursor variables do together. Of course, operational use of this latter model unrealistically requires perfect forecasts for the Niño-3.4 SST. Again we conclude that ENSO predictability is currently the missing requirement for more skillful rainfall forecasts at longer lead times.

The CCA defines spatial pattern relationships between global SST and JJAS Ethiopia station rainfalls. The simultaneous SST–rainfall patterns strongly confirm the impact of ENSO, and indicate a lesser role for SSTs near the source regions of monsoonal low-level systems near southwest India and in South Atlantic. These conclusions also apply to the CCA using leading May SSTs.

In summary, this study’s main finding is that the northern summer ENSO condition is overwhelmingly the single most important factor governing the JJAS rainfall across Ethiopia, excluding the southern/southeastern lowlands. SST anomalies in the Atlantic and Indian Oceans appear to matter far less. More regional climate and weather processes were not investigated here, but could be tied into this larger scale. Skillful predictions of Ethiopian summer rainfall hinge upon the best possible forecasts of the summer ENSO state from an earlier time. Useful summer rainfall predictions are thus potentially achievable using global dynamical or statistical models. Further study may extend knowledge to more regional scales, particularly using regional models that reproduce large-scale processes (e.g., ITCZ, middle- and upper-troposphere circulations), and downscale for local land surface variations. In the meantime, existing statistical modeling techniques, aided by statistical or dynamical predictions of the summer ENSO state, allow for improved use of seasonal rainfall forecasts for sustainable, dependable early warning systems so critically important to societal operations.

Acknowledgments

The authors appreciate the helpful comments of the anonymous reviewers, as well as those of Z. T. Segele and W. Thiaw. Z. Segale kindly offered Fig. 13 to us. We are grateful to Mark Cane for directing the Climate and Society masters program that enabled the Korecha–Barnston collaboration underlying this study, to Dake Chen for a special hindcast run of the LDEO-5 ENSO forecast model, and to Simon Mason for developing the CPT, which contained the user-friendly CCA package used in this study. This paper is funded by a Cooperative Agreement NA050AR431004 from the National Oceanic and Atmospheric Administration. The views expressed herein are those of the authors and do not necessarily reflect the views of NOAA or any of its subagencies.

REFERENCES

REFERENCES
Barnett
,
T. P.
, and
R.
Preisendorfer
,
1987
:
Origins and levels of monthly and seasonal forecast skill for U.S. surface air temperatures determined by canonical correlation analysis.
Mon. Wea. Rev.
,
115
,
1825
1850
.
Barnston
,
A. G.
, and
T. M.
Smith
,
1996
:
Specification and prediction of global surface temperature and precipitation from global SST using CCA.
J. Climate
,
9
,
2660
2697
.
Barnston
,
A. G.
, and
Coauthors
,
1994
:
Long-lead seasonal forecasts—Where do we stand?
Bull. Amer. Meteor. Soc.
,
75
,
2097
2114
.
Barnston
,
A. G.
,
M.
Chelliah
, and
S. B.
Goldenberg
,
1997
:
Documentation of a highly ENSO-related SST region in the equatorial Pacific.
Atmos.–Ocean
,
35
,
367
383
.
Bekele
,
F.
,
1997
:
Ethiopian use of ENSO information in its seasonal forecasts. Internet J. Afr. Studies, No. 2. [Available online at http://www.ccb.ucar.edu/ijas/ijasno2/bekele.html.]
.
Bhatt
,
U. S.
,
1989
:
Circulation regimes of rainfall anomalies in the African–South Asian Monsoon Belt.
J. Climate
,
2
,
1133
1144
.
Camberlin
,
P.
,
1997
:
Rainfall anomalies in the source region of the Nile and their connection with the Indian summer monsoon.
J. Climate
,
10
,
1380
1392
.
Cane
,
M. A.
,
2000
:
Understanding and predicting the world’s climate system.
Applications of Seasonal Climate Forecasting in Agricultural and Natural Ecosystems—The Australian Experience, G. Hammer, Ed., Kluwer Academic Publishers, 29–50
.
Cane
,
M. A.
, and
S. E.
Zebiak
,
1987
:
Deterministic prediction of El Niño events.
Atmosphere and Oceanic Variability, H. Cattle, Ed., Roy. Meteor. Soc., 153–182
.
Cane
,
M. A.
,
S. E.
Zebiak
, and
S. C.
Dolan
,
1986
:
Experimental forecasts for El-Niño.
Nature
,
321
,
827
832
.
Chen
,
D.
,
M. A.
Cane
,
A.
Kaplan
, and
S. E.
Zebiak
,
2004
:
Predictability of El Niño over the past 148 years.
Nature
,
428
,
733
735
.
Cook
,
K. H.
,
1997
:
Large-scale atmospheric dynamics and Sahelian precipitation.
J. Climate
,
10
,
1137
1152
.
Degefu
,
W.
,
1987
:
Some aspects of meteorological drought in Ethiopia.
Drought and Hunger in Africa: Denying Famine a Future, M. H. Glantz, Ed., Cambridge University Press, 23–36
.
Folland
,
C.
,
J.
Owen
,
M. N.
Ward
, and
A.
Colman
,
1991
:
Prediction of seasonal rainfall in the Sahel region using empirical and dynamical models.
J. Forecasting
,
10
,
21
56
.
Fraedrich
,
K.
,
J.
Jiang
,
F-W.
Gerstengarbe
, and
P. C.
Werner
,
1997
:
Multiscale detection of abrupt climate changes: Application to River Nile flood levels.
Int. J. Climatol.
,
17
,
1301
1315
.
Giannini
,
A.
,
R.
Saravana
, and
P.
Chang
,
2003
:
Oceanic forcing of Sahel rainfall on inter-annual to inter-decadal time scales.
Science
,
302
,
1027
1030
.
Gissila
,
T.
,
E.
Black
,
D. I. F.
Grimes
, and
J. M.
Slingo
,
2004
:
Seasonal forecasting of the Ethiopian summer rains.
Int. J. Climatol.
,
24
,
1345
1358
.
Glahn
,
H. R.
,
1968
:
Canonical correlation and its relationship to discriminant analysis and multiple regression.
J. Atmos. Sci.
,
25
,
23
31
.
Goddard
,
L.
, and
N. E.
Graham
,
1999
:
Importance of the Indian Ocean for simulating rainfall anomalies over eastern and southern Africa.
J. Geophys. Res.
,
104
,
19099
19116
.
Goddard
,
L.
, and
M.
Dilley
,
2005
:
El Niño: Catastrophe or opportunity.
J. Climate
,
18
,
651
665
.
Goddard
,
L.
,
S. J.
Mason
,
S. E.
Zebiak
,
R. E.
Ropelewski
,
R.
Basher
, and
M. A.
Cane
,
2001
:
Current approaches to seasonal-to-interannual climate predictions.
Int. J. Climatol.
,
21
,
1111
1152
.
Goldammer
,
J. G.
,
1999
:
Forests on fire.
Science
,
284
,
1782
1783
.
Gong
,
X.
,
A. G.
Barnston
, and
M. N.
Ward
,
2003
:
The effect of spatial aggregation on the skill of seasonal precipitation forecasts.
J. Climate
,
16
,
3059
3071
.
Grist
,
J. P.
,
E. S.
Nicholson
, and
A. I.
Barcilon
,
2002
:
Easterly waves over Africa. Part II: Observed and modeled contrasts between wet and dry years.
Mon. Wea. Rev.
,
130
,
212
225
.
Hastenrath
,
S.
,
1991
:
Climate Dynamics of the Tropics.
Kluwer Academic Publishers, 488 pp
.
Hastenrath
,
S.
,
1995
:
Recent advances in tropical climate prediction.
J. Climate
,
8
,
1519
1532
.
Hastenrath
,
S.
,
2000
:
Interannual and longer term variability of upper-air circulation over the tropical Atlantic and West Africa in boreal summer.
Int. J. Climatol.
,
20
,
1415
1430
.
Hotelling
,
H.
,
1936
:
Relations between two sets of variates.
Biometrika
,
28
,
321
377
.
Hulme
,
M.
,
2001
:
Climate perspectives on Sahelian desiccation: 1973–1998.
Global Environ. Change
,
11
,
19
29
.
Indeje
,
M. H.
,
M. F.
Semazzi
, and
L. J.
Ogallo
,
2000
:
ENSO signals in East African rainfall seasons.
Int. J. Climatol.
,
20
,
19
46
.
Janicot
,
S.
,
V.
Moron
, and
B.
Fontaine
,
1996
:
Sahel droughts and ENSO dynamics.
Geophys. Res. Lett.
,
23
,
515
518
.
Janicot
,
S.
,
S.
Trzaska
, and
I.
Poccard
,
2001
:
Summer Sahel–ENSO teleconnection and decadal time scale SST variations.
Climate Dyn.
,
18
,
303
320
.
Jury
,
M. R.
,
2002
:
Economic impacts of climate variability in South Africa and development of resource prediction models.
J. Appl. Meteor.
,
41
,
46
55
.
Kassahun
,
B.
,
1987
:
Weather systems over Ethiopia. Proc. First Tech. Conf. on Meteorological Research in Eastern and Southern Africa, Nairobi, Kenya, UCAR, 53–57
.
Khandekar
,
M. L.
,
T. S.
Murty
,
D.
Scott
, and
W.
Baird
,
2000
:
The 1997 El Niño, Indonesian forest fires and the Malaysian smoke problem: A deadly combination of natural and man-made hazards.
Nat. Hazards
,
21
,
131
144
.
Kidson
,
J. W.
,
1977
:
African rainfall and its relation to the upper air circulation.
Quart. J. Roy. Meteor. Soc.
,
103
,
441
456
.
Lamb
,
H. H.
,
1977
:
Some comments on the drought in recent years in the Sahel–Ethiopia zone of North Africa.
Drought in Africa, D. Dalby, R. J. Harrison-Church, and F. Bazzaz, Eds., Vol. 2, School of Oriental and African Studies, 33–37
.
Lamb
,
H. H.
,
1978
:
Large-scale tropical Atlantic surface circulation patterns associated with Sub-Saharan weather.
Tellus
,
30
,
240
251
.
Latif
,
M.
, and
Coauthors
,
1998
:
A review of the predictability and prediction of ENSO.
J. Geophys. Res.
,
103
,
14375
14393
.
Mason
,
S. J.
, and
L.
Goddard
,
2001
:
Probabilistic precipitation anomalies associated with ENSO.
Bull. Amer. Meteor. Soc.
,
82
,
619
638
.
Michaelsen
,
J.
,
1987
:
Cross-validation in statistical climate forecast models.
J. Climate Appl. Meteor.
,
26
,
1589
1600
.
Mutai
,
C. C.
,
M. N.
Ward
, and
A. W.
Colman
,
1998
:
Towards the prediction of the East Africa short rains based on sea-surface temperature-atmosphere coupling.
Int. J. Climatol.
,
18
,
975
997
.
Nicholls
,
N.
,
1993
:
What are the potential contributions of El Niño–Southern Oscillation research to early warning of potential acute food-deficit situations?
Workshop on Usable Science: Food Security, Early Warning and El Niño, M. H. Glantz, Ed., National Center for Atmospheric Research, 169–177
.
Nicholson
,
S. E.
,
1989
:
Long-term changes in African rainfall.
Weather
,
44
,
46
56
.
Nicholson
,
S. E.
, and
J. Y.
Kim
,
1997
:
The relationship of the El Niño–Southern Oscillation to African rainfall.
Int. J. Climatol.
,
17
,
117
135
.
Nicholson
,
S. E.
, and
J. C.
Selato
,
2000
:
The influence of La Niña on African rainfall.
Int. J. Climatol.
,
20
,
1761
1776
.
NMSA
,
1996
:
Climatic and agroclimatic resources of Ethiopia. National Meteorological Services Agency of Ethiopia, Meteorological Research Report Series, Vol. 1, No. 1, 1–137
.
Ntale
,
H. R.
,
T. Y.
Gan
, and
D.
Mwale
,
2003
:
Prediction of East African seasonal rainfall using simplex canonical correlation analysis.
J. Climate
,
16
,
2105
2112
.
Ogallo
,
L. J.
,
1988
:
Relationships between seasonal rainfall in East Africa and the Southern Oscillation.
Int. J. Climatol.
,
8
,
31
43
.
Ogallo
,
L. J.
,
1989
:
The spatial and temporal patterns of the East African seasonal rainfall derived from principal component analysis.
Int. J. Climatol.
,
9
,
145
167
.
Osman
,
Y. Z.
, and
Y. A.
Shamseldin
,
2002
:
Quantitative rainfall prediction models for Central and Southern Sudan using El Niño–Southern Oscillation and Indian Ocean sea surface temperature indices.
Int. J. Climatol.
,
22
,
1861
1878
.
Paeth
,
H.
, and
P.
Friederichs
,
2004
:
Seasonality and time scales in the relationship between global SST and African rainfall.
Climate Dyn.
,
23
,
815
837
.
Philippon
,
N. P.
,
P.
Camberin
, and
N.
Fauchereau
,
2002
:
Empirical predictability study of October–December East African rainfall.
Quart. J. Roy. Meteor. Soc.
,
128
,
2239
2256
.
Raicich
,
F.
,
N.
Pinardi
, and
A.
Navarra
,
2003
:
Teleconnections between Indian monsoon and Sahel rainfall and the Mediterranean.
Int. J. Climatol.
,
23
,
173
186
.
Ramsey
,
F. L.
, and
W. D.
Schafer
,
2002
:
The Statistical Sleuth: A Course in Methods of Data Analysis.
2d ed. Duxbury, 742 pp
.
Ropelewski
,
C. F.
, and
M. S.
Halpert
,
1987
:
Global- and regional scale precipitation patterns associated with the El Niño–Southern Oscillation.
Mon. Wea. Rev.
,
115
,
1606
1626
.
Rousseeuw
,
P. J.
, and
A. M.
Leroy
,
1987
:
Robust Regression and Outlier Detection.
John Wiley & Sons, Inc., 360 pp
.
Rowell
,
D. P.
,
2001
:
Teleconnection between the tropical Pacific and the Sahel.
Quart. J. Roy. Meteor. Soc.
,
127
,
1683
1706
.
Rowell
,
D. P.
,
2003
:
The impact of Mediterranean SSTs on the Sahelian rainfall season.
J. Climate
,
16
,
849
862
.
Rowell
,
D. P.
,
C. K.
Folland
,
K.
Maskell
, and
M. N.
Ward
,
1995
:
Variability of summer rainfall over tropical North Africa (1906–92): Observations and modeling.
Quart. J. Roy. Meteor. Soc.
,
121
,
669
704
.
Segele
,
Z. T.
, and
P. J.
Lamb
,
2005
:
Characterization and variability of Kiremt rainy season over Ethiopia.
Meteor. Atmos. Phys.
,
89
,
153
180
.
Shanko
,
D.
, and
P.
Camberlin
,
1998
:
The effect of the southwest Indian Ocean tropical cyclones on Ethiopian drought.
Int. J. Climatol.
,
18
,
1373
1378
.
Sileshi
,
Y.
, and
G. R.
Demarée
,
1995
:
Rainfall variability in the Ethiopian and Eritrean highlands and its links with the Southern Oscillation Index.
J. Biogeogr.
,
22
,
945
952
.
Smith
,
T. M.
, and
R. W.
Reynolds
,
2004
:
Improved extended reconstruction of SST (1854–1997).
J. Climate
,
17
,
2466
2477
.
Tadesse
,
T.
,
1994
:
The influence of the Arabian Sea storms/depressions over the Ethiopian weather. Proc. Int. Conf. on Monsoon Variability and Prediction, WCRP-84 and WMO Tech. Doc. 619, Geneva, Switzerland, World Meteorological Organization, 228–236
.
Thiaw
,
W. M.
,
A. G.
Barnston
, and
V.
Kumar
,
1999
:
Predictions of African rainfall on the seasonal timescale.
J. Geophys. Res.
,
104
,
31589
31597
.
Trenberth
,
K. E.
, and
T. J.
Hoar
,
1996
:
The 1990–1995 El Niño–Southern Oscillation event: Longest on record.
Geophys. Res. Lett.
,
23
,
57
60
.
Tsegay
,
W.
,
1998
:
El Niño and drought early warning in Ethiopia.
Internet J. Afr. Studies
, .
Tsegay
,
W.
,
2001
:
The case of Ethiopia: Impacts and responses to the 1997-98 El Niño event.
Once Burned, Twice Shy? Lessons Learned from the 1997-98 El Niño, M. H. Glantz, Ed., United Nations University Press, 88–100
.
Tziperman
,
E.
,
M. A.
Cane
,
S. E.
Zebiak
,
Y.
Xue
, and
B.
Blumenthal
,
1998
:
Locking of El Niño’s peak time to the end of the calendar year in the delayed oscillator picture of ENSO.
J. Climate
,
11
,
2191
2199
.
Wang
,
G.
,
E. A. B.
Eltahir
,
J. A.
Foley
, and
D.
Pollard
,
2004
:
Decadal variability of rainfall in the Sahel: Results from the coupled GENESIS-IBIS atmosphere–biosphere model.
Climate Dyn.
,
22
,
625
637
.
Ward
,
M. N.
,
1998
:
Diagnosis and short-lead time prediction of summer rainfall in tropical North Africa at interannual and multidecadal time scales.
J. Climate
,
11
,
3167
3191
.
Wright
,
P. B.
,
J. M.
Wallace
,
T. P.
Mitchell
, and
C.
Deser
,
1988
:
Correlation structure of the El Niño–Southern Oscillation phenomenon.
J. Climate
,
1
,
609
625
.
Zebiak
,
S. E.
, and
M. A.
Cane
,
1987
:
A model El Niño–Southern Oscillation.
Mon. Wea. Rev.
,
115
,
2262
2278
.
Zeng
,
N.
,
2003
:
Drought in the Sahel.
Science
,
302
,
999
1000
.
Zeng
,
N.
,
J. D.
Neelin
, and
C. J.
Tucker
,
1999
:
Enhancement of interdecadal climate variability in the Sahel by vegetation interaction.
Science
,
286
,
1537
1540
.

Footnotes

Corresponding author address: Anthony G. Barnston, International Research Institute for Climate and Society, 61 Route 9W, Palisades, NY 10962. Email: tonyb@iri.columbia.edu

1

Note that droughts and floods are sometimes declared even when total rainfall or the number of days receiving rain is not very anomalous. This can occur when seasonal rainfall is grossly unevenly distributed over the season, with long dry or wet spells that may straddle the monthly boundaries.

2

A few of the analyses use data spanning only through 2000 or 2002, owing to practical considerations, and this is always noted when it is the case.

3

Robust regression enables assessment of the sensitivity of results to outlier cases, if any are present.

4

Climate prediction in Ethiopia started in 1987 as an experimental innovation (NMSA 1996) after Cane and Zebiak (1987) introduced the first ENSO prediction model. This beginning was a result of research undertaken at NMSA in Ethiopia (Degefu 1987; NMSA 1996). The widespread Ethiopian famines in 1972/73, 1982/83 and 1984/85, confirmed to be drought related, could have been greatly diminished using early warning systems based on probabilistic seasonal rainfall predictions. In the early stages of seasonal prediction, regional synoptic patterns were emphasized (Kassahun 1987; Tadesse 1994). However, the coinciding of drought years with El Niño attracted attention to ENSO as a vital predictor, as anticipating the ENSO state prior to the summer could sharpen the rainfall predictions. In the late 1980s and early 1990s, NMSA added new tools to its initially synoptically based seasonal rainfall predictions such as analogs and ENSO teleconnections, and achieved favorable results in anticipating some drought and flood catastrophes. In summer of 1995 the analog tool was augmented to include ENSO, Atlantic and Indian Ocean SSTs, and large-scale regional circulation patterns. The two to three best analog years were used to suggest the tercile-based seasonal rainfall probabilities as well as the character of the intraseasonal variability, both applied to individual Ethiopian regions.

5

NOAA defines a nonneutral ENSO state as a departure from normal of the SST in the Niño-3.4 region of magnitude 0.5°C or more, lasting for at least five running three-month periods.

6

It is worth noting that when the rainfall index is computed standardizing at the station level before doing so again for the 55-station average, this correlation is −0.77.

7

The low rainfall for 1987 is identified as an influential point (Ramsey and Schafer 2002). Four robust regression alternatives (least trimmed squares regression, least median squares regression, least absolute deviations regression, and maximum likelihood estimates of regression) were applied (Rousseeuw and Leroy 1987) with and without 1987, and comparison of results of the predictor coefficients with those of the original ordinary least squares model indicate that the original model performs entirely satisfactorily without special outlier accommodation.

8

A CCA software module called Climate Predictability Tool (CPT) was downloaded from a Web page of the International Research Institute for Climate and Society (IRI) and applied to the global SST and rainfall data used here.