• Ba, M. B., , and Nicholson S. E. , 1998: Analysis of convective activity and its relationship to the rainfall over the Rift Valley lakes of East Africa during 1983–90 using the Meteosat infrared channel. J. Appl. Meteor., 37, 12501264, doi:10.1175/1520-0450(1998)037<1250:AOCAAI>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Batte, L., , and Deque M. , 2011: Seasonal predictions of precipitation over Africa using coupled ocean-atmosphere general circulation models: Skill of the ENSEMBLES project multimodel ensemble forecasts. Tellus, 63A, 283299, doi:10.1111/j.1600-0870.2010.00493.x.

    • Search Google Scholar
    • Export Citation
  • Block, P., , and Rajagopalan B. , 2007: Interannual variability and ensemble forecast of upper Blue Nile basin kiremt season precipitation. J. Hydrometeor., 8, 327343, doi:10.1175/JHM580.1.

    • Search Google Scholar
    • Export Citation
  • Camberlin, P., 1995: June–September rainfall in north-eastern Africa and atmospheric signals over the tropics: A zonal perspective. Int. J. Climatol., 15, 773783, doi:10.1002/joc.3370150705.

    • Search Google Scholar
    • Export Citation
  • Camberlin, P., , and Planchon O. , 1997: Coastal precipitation regimes in Kenya. Geogr. Ann., 79A, 109119, doi:10.1111/1468-0459.00010.

  • Camberlin, P., , and Philippon N. , 2002: The East African March–May rainy season: Associated atmospheric dynamics and predictability over the 1968–97 period. J. Climate, 15, 10021019, doi:10.1175/1520-0442(2002)015<1002:TEAMMR>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Dee, D. P., and et al. , 2011: The ERA-Interim reanalysis: Configuration and performance of the data assimilation system. Quart. J. Roy. Meteor. Soc., 137, 553597, doi:10.1002/qj.828.

    • Search Google Scholar
    • Export Citation
  • Dezfuli, A. K., , and Nicholson S. E. , 2013: The relationship of interannual variability in western equatorial Africa to the tropical oceans and atmospheric circulation. Part II: The boreal autumn. J. Climate, 26, 6684, doi:10.1175/JCLI-D-11-00686.1.

    • Search Google Scholar
    • Export Citation
  • Diro, G. T., , Grimes D. I. F. , , and Black E. , 2008: Seasonal forecasting of Ethiopian spring rains. Meteor. Appl., 15, 7383, doi:10.1002/met.63.

    • Search Google Scholar
    • Export Citation
  • Diro, G. T., , Grimes D. I. F. , , and Black E. , 2011a: Teleconnections between Ethiopian summer rainfall and sea surface temperature: Part I—Observation and modelling. Climate Dyn., 37, 103119, doi:10.1007/s00382-010-0837-8.

    • Search Google Scholar
    • Export Citation
  • Diro, G. T., , Grimes D. I. F. , , and Black E. , 2011b: Teleconnections between Ethiopian summer rainfall and sea surface temperature: Part II: Seasonal forecasting. Climate Dyn., 37, 121131, doi:10.1007/s00382-010-0896-x.

    • Search Google Scholar
    • Export Citation
  • Diro, G. T., , Tompkins A. M. , , and Bi X. , 2012: Dynamical downscaling of ECMWF Ensemble seasonal forecasts over East Africa with RegCM3. J. Geophys. Res.,117, D16103, doi:10.1029/2011JD016997.

  • Elsner, J. B., , and Schmertmann C. P. , 1994: Assessing forecast skill through cross-validation. Wea. Forecasting, 9, 619624, doi:10.1175/1520-0434(1994)009<0619:AFSTCV>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Friederichs, P., , and Paeth H. , 2006: Seasonal prediction of African prediction with ECHAM4-T42 ensemble simulations using a multivariate MOS prediction scheme. Climate Dyn., 27, 761786, doi:10.1007/s00382-006-0154-4.

    • Search Google Scholar
    • Export Citation
  • Gissila, T., , Black E. , , Grimes D. , , and Slingo J. , 2004: Seasonal forecasting of the Ethiopian summer rains. Int. J. Climatol., 24, 13451358, doi:10.1002/joc.1078.

    • Search Google Scholar
    • Export Citation
  • Hastenrath, S., , and Polzin D. , 2005: Mechanisms of climate anomalies in the equatorial Indian Ocean. J. Geophys. Res.,110, D0811, doi:10.1029/2004JD004981.

    • Search Google Scholar
    • Export Citation
  • Hastenrath, S., , Nicklis A. , , and Greischer L. , 1993: Atmospheric-hydrospheric mechanisms of climate anomalies in the western equatorial Indian Ocean. J. Geophys. Res., 98, 20 219–20 235, doi:10.1029/93JC02330.

    • Search Google Scholar
    • Export Citation
  • Hastenrath, S., , Polzin D. , , and Camberlin P. , 2004: Exploring the predictability of the ‘short rains’ at the coast of East Africa. Int. J. Climatol., 24, 13331343, doi:10.1002/joc.1070.

    • Search Google Scholar
    • Export Citation
  • Kalnay, E., and et al. , 1996: The NCEP/NCAR 40-Year Reanalysis Project. Bull. Amer. Meteor. Soc., 77, 437471, doi:10.1175/1520-0477(1996)077<0437:TNYRP>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Kistler, R., and et al. , 2001: The NCEP–NCAR 50-Year Reanalysis: Monthly means CD-ROM and documentation. Bull. Amer. Meteor. Soc., 82, 247267, doi:10.1175/1520-0477(2001)082<0247:TNNYRM>2.3.CO;2.

    • Search Google Scholar
    • Export Citation
  • Korecha, D., , and Barnston A. G. , 2007: Predictability of June–September rainfall in Ethiopia. Mon. Wea. Rev., 135, 628650, doi:10.1175/MWR3304.1.

    • Search Google Scholar
    • Export Citation
  • Lamb, P. J., , and Peppler R. A. , 1992: Further case studies of tropical Atlantic surface atmospheric and oceanic patterns associated with sub-Saharan drought. J. Climate, 5, 476488, doi:10.1175/1520-0442(1992)005<0476:FCSOTA>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Lebel, T., , and Ali A. , 2009: Recent trends in the Central and western Sahel rainfall regime (1990–2007). J. Hydrol., 375, 5264, doi:10.1016/j.jhydrol.2008.11.030.

    • Search Google Scholar
    • Export Citation
  • Lyon, B., , and DeWitt D. G. , 2012: A recent and abrupt decline in the East African long rains. Geophys. Res. Lett., 39, L02702, doi:10.1029/2011GL050337.

    • Search Google Scholar
    • Export Citation
  • Michaelsen, J., 1987: Cross-validation in statistical climate forecast models. J. Climate Appl. Meteor., 26, 15891600, doi:10.1175/1520-0450(1987)026<1589:CVISCF>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Mutai, C. C., , Ward M. N. , , and Colman A. W. , 1998: Towards the prediction of the East Africa short rains based on sea-surface temperature–atmosphere coupling. Int. J. Climatol., 18, 975997, doi:10.1002/(SICI)1097-0088(199807)18:9<975::AID-JOC259>3.0.CO;2-U.

    • Search Google Scholar
    • Export Citation
  • Mwale, D., , and Gan T. Y. , 2005: Wavelet analysis of variability, teleconnectivity, and predictability of the September–November East African rainfall. J. Appl. Meteor., 44, 256269, doi:10.1175/JAM2195.1.

    • Search Google Scholar
    • Export Citation
  • Ndiaye, O., , Goddard L. , , and Ward M. N. , 2009: Using regional wind fields to improve general circulation model forecasts of July–September Sahel rainfall. Int. J. Climatol., 29, 12621275, doi:10.1002/joc.1767.

    • Search Google Scholar
    • Export Citation
  • Nicholson, S. E., 1986: The spatial coherence of African rainfall anomalies: Interhemispheric teleconnections. J. Climate Appl. Meteor., 25, 13651381, doi:10.1175/1520-0450(1986)025<1365:TSCOAR>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Nicholson, S. E., 1993: An overview of African rainfall fluctuations of the last decade. J. Climate, 6, 14631466, doi:10.1175/1520-0442(1993)006<1463:AOOARF>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Nicholson, S. E., 1996: A review of climate dynamics and climate variability in eastern Africa. The Limnology, Climatology and Paleoclimatology of the East African Lakes, T. C. Johnson and E. Odada, Eds., Gordon and Breach, 25–56.

  • Nicholson, S. E., 1997: An analysis of the ENSO signal in the tropical Atlantic and western Indian Oceans. Int. J. Climatol., 17, 345375, doi:10.1002/(SICI)1097-0088(19970330)17:4<345::AID-JOC127>3.0.CO;2-3.

    • Search Google Scholar
    • Export Citation
  • Nicholson, S. E., 2000: The nature of rainfall variability over Africa on time scales of decades to millennia. Global Planet. Change, 26, 137158, doi:10.1016/S0921-8181(00)00040-0.

    • Search Google Scholar
    • Export Citation
  • Nicholson, S. E., 2013: A detailed look the recent drought situation in the Great Horn of Africa. J. Arid Environ., 103, 71–79, doi:10.1016/j.jaridenv.2013.12.003.

  • Nicholson, S. E., , and Dezfuli A. K. , 2013: The relationship of interannual variability in western equatorial Africa to the tropical oceans and atmospheric circulation. Part I: The boreal spring. J. Climate, 26, 4565, doi:10.1175/JCLI-D-11-00653.1.

    • Search Google Scholar
    • Export Citation
  • Nicholson, S. E., , Some B. , , and Kone B. , 2000: An analysis of recent rainfall conditions in West Africa including the rainy seasons of the 1997 El Niño and the 1998 La Niña years. J. Climate, 13, 26282640, doi:10.1175/1520-0442(2000)013<2628:AAORRC>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Ntale, H. K., , Gan T. Y. , , and Mwale D. , 2003: Prediction of East African seasonal rainfall using simplex canonical correlation analysis. J. Climate, 16, 21052112, doi:10.1175/1520-0442(2003)016<2105:POEASR>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Philippon, N., , Camberlin P. , , and Fauchereau N. , 2002: Empirical predictability study of October–December East African rainfall. Quart. J. Roy. Meteor. Soc., 128, 22392256, doi:10.1256/qj.01.190.

    • Search Google Scholar
    • Export Citation
  • Segele, Z. T., , Lamb P. J. , , and Leslie L. M. , 2009a: Seasonal-to-interannual variability of Ethiopia/Horn of Africa monsoon. Part I: Associations of wavelet-filtered large-scale atmospheric circulation and global sea surface temperature. J. Climate, 22, 33963421, doi:10.1175/2008JCLI2859.1.

    • Search Google Scholar
    • Export Citation
  • Segele, Z. T., , Lamb P. J. , , and Leslie L. M. , 2009b: Large-scale atmospheric circulation and global sea surface temperature associations with Horn of Africa June–September rainfall. Int. J. Climatol., 29, 10751100, doi:10.1002/joc.1751.

    • Search Google Scholar
    • Export Citation
  • Smith, T. M., , Reynolds R. W. , , Peterson T. C. , , and Lawrimore J. , 2008: Improvements to NOAA’s historical merged land–ocean surface temperature analysis (1880–2006). J. Climate, 21, 22832296, doi:10.1175/2007JCLI2100.1.

    • Search Google Scholar
    • Export Citation
  • Thiaw, W. M., , Barnston A. G. , , and Kumar V. , 1999: Predictions of African rainfall on the seasonal timescale. J. Geophys. Res., 104, 31 58931 597, doi:10.1029/1999JD900906.

    • Search Google Scholar
    • Export Citation
  • Torrence, C., , and Webster P. J. , 1998: The annual cycle of persistence in the El Niño/Southern Oscillation. Quart. J. Roy. Meteor. Soc., 124, 19852004, doi:10.1002/qj.49712455010.

    • Search Google Scholar
    • Export Citation
  • Tsidu, M., 2012: High-resolution monthly rainfall database for Ethiopia: Homogenization, reconstruction, and gridding. J. Climate, 25, 84228443, doi:10.1175/JCLI-D-12-00027.1.

    • Search Google Scholar
    • Export Citation
  • van Oldenborgh, G. J., , Balmaseda M. , , Ferranta L. , , Stockdale T. , , and Anderson D. , 2005: Did the ECMWF seasonal forecast model outperform statistical ENSO forecast models over the last 15 years? J. Climate, 18, 32403249, doi:10.1175/JCLI3420.1.

    • Search Google Scholar
    • Export Citation
  • Wajsowicz, R. C., 2007: Seasonal-to-interannual forecasting of tropical Indian Ocean sea surface temperature anomalies: Potential predictability and barriers. J. Climate, 20, 33203343, doi:10.1175/JCLI4162.1.

    • Search Google Scholar
    • Export Citation
  • Webster, P. J., , and Yang S. , 1992: Monsoon and ENSO: Selectively interactive systems. Quart. J. Roy. Meteor. Soc., 118, 877926, doi:10.1002/qj.49711850705.

    • Search Google Scholar
    • Export Citation
  • Webster, P. J., , Moore A. , , Loschnigg J. , , and Leban M. , 1999: Coupled ocean–atmosphere dynamics in the Indian Ocean during 1997–98. Nature, 401, 356360, doi:10.1038/43848.

    • Search Google Scholar
    • Export Citation
  • Williams, A. P., and et al. , 2012: Recent summer precipitation trends in the Greater Horn of Africa and the emerging role of Indian Ocean sea surface temperature. Climate Dyn., 39, 23072328, doi:10.1007/s00382-011-1222-y.

    • Search Google Scholar
    • Export Citation
  • View in gallery

    (top) Month of rainfall maximum and area rainfall divisions (defined by numbers with circles 1–4). (bottom four panels) Monthly rainfall (mm) at typical stations: Arba Minch, Addis Ababa, Embu, and Migori.

  • View in gallery

    Mean monthly rainfall (mm) over the study area.

  • View in gallery

    Mean vector winds by season for (top) 925, (middle) 700, and (bottom) 200 mb: (left) MAM, (middle) JAS, and (right) ON.

  • View in gallery

    Low-level (925 mb) divergence in October (10−6 s−1).

  • View in gallery

    Stations used in the rainfall analysis with the black dots identifying a summer rainfall region and blue, an equatorial one.

  • View in gallery

    Location of predictors (colors, see Table 1) used for the MAM season: (top) equatorial rainfall region and (bottom) summer rainfall region.

  • View in gallery

    Predicted (red) vs observed (black) MAM rainfall for the equatorial region based on (top) February and (middle) January predictors; and (bottom) for the summer rainfall region based on February predictors. Correlation r between the predicted and observed is indicated in each panel.

  • View in gallery

    Location of predictors (colors, see Table 2) used for the JAS season for the equatorial and summer rainfall regions.

  • View in gallery

    Predicted (red) vs observed (black) JAS rainfall based on May predictors: (top) equatorial and (bottom) summer regions. Correlation r between the predicted and observed is indicated in each panel.

  • View in gallery

    Location of predictors (Table 3) used for the ON season: (top) equatorial rainfall and (bottom) summer rainfall regions. Red lines indicate that the predictor is the difference between the variable in the two liked boxes.

  • View in gallery

    Predicted (red) vs observed (black) ON rainfall for the (top two panels) equatorial and (bottom two panels) summer rainfall regions with August predictors in one case (first and third panels) and May predictors in a second case (second and fourth panels). Correlation r between the predicted and observed is indicated in each panel.

  • View in gallery

    Example of cross validation for (top) a case with high correlation between predicted (red) and observed (black) (summer rainfall region, JAS rainfall, and May predictors) and (bottom) a case with a low correlation between predicted and observed (equatorial rainfall region, JAS rainfall, and May predictors). Correlation r between the predicted and observed is indicated in each panel.

  • View in gallery

    Predicted (red) vs observed (black) rainfall for a subregion in the Kenyan highlands. (top) ON rainfall based on August predictors. (bottom) MAM rainfall based on January predictors. Correlation r between the predicted and observed is indicated in each panel.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 130 130 26
PDF Downloads 133 133 41

The Predictability of Rainfall over the Greater Horn of Africa. Part I: Prediction of Seasonal Rainfall

View More View Less
  • 1 Department of Earth, Ocean and Atmospheric Science, Florida State University, Tallahassee, Florida
© Get Permissions
Full access

Abstract

The predictability of each of the three rainy seasons affecting the Horn of Africa is examined using multiple linear regression and cross validation. In contrast to most other empirical forecast models, atmospheric dynamics are emphasized. Two geographical sectors are considered: the summer rainfall region with a single rainfall peak in the boreal summer and an equatorial rainfall region with rainy seasons in both the boreal spring and the boreal autumn. These two seasons are termed the “long rains” and “short rains,” respectively, in much of East Africa. Excellent predictability is noted 5 months in advance for both regions during the boreal autumn season and 2 months in advance for the summer season in the summer rainfall region. There is also excellent predictability for the short rains of the equatorial region 2 months in advance. Two notable findings are that atmospheric variables generally provide higher forecast skill than surface variables, such as sea surface temperatures and sea level pressure, and that ENSO and the Indian Ocean dipole provide less forecast skill than atmospheric variables associated with them. As in other studies, the results show that the spring predictability barrier limits the lead time for the forecasting of spring and summer rainfall.

Corresponding author address: Sharon Nicholson, Department of Earth, Ocean and Atmospheric Science, Florida State University, Love Building, Tallahassee, FL 32306. E-mail: snicholson@fsu.edu

Abstract

The predictability of each of the three rainy seasons affecting the Horn of Africa is examined using multiple linear regression and cross validation. In contrast to most other empirical forecast models, atmospheric dynamics are emphasized. Two geographical sectors are considered: the summer rainfall region with a single rainfall peak in the boreal summer and an equatorial rainfall region with rainy seasons in both the boreal spring and the boreal autumn. These two seasons are termed the “long rains” and “short rains,” respectively, in much of East Africa. Excellent predictability is noted 5 months in advance for both regions during the boreal autumn season and 2 months in advance for the summer season in the summer rainfall region. There is also excellent predictability for the short rains of the equatorial region 2 months in advance. Two notable findings are that atmospheric variables generally provide higher forecast skill than surface variables, such as sea surface temperatures and sea level pressure, and that ENSO and the Indian Ocean dipole provide less forecast skill than atmospheric variables associated with them. As in other studies, the results show that the spring predictability barrier limits the lead time for the forecasting of spring and summer rainfall.

Corresponding author address: Sharon Nicholson, Department of Earth, Ocean and Atmospheric Science, Florida State University, Love Building, Tallahassee, FL 32306. E-mail: snicholson@fsu.edu

1. Introduction

The Horn of Africa (Fig. 1) is noted for the extreme interannual variability of its rainfall regime. In several recent cases, drought years have alternated with flood years. During the year extending from July 2010 to June 2011, rainfall was at least 50%–75% below normal in roughly half of the region. Record floods followed in October 2011 (Nicholson 2013).

Fig. 1.
Fig. 1.

(top) Month of rainfall maximum and area rainfall divisions (defined by numbers with circles 1–4). (bottom four panels) Monthly rainfall (mm) at typical stations: Arba Minch, Addis Ababa, Embu, and Migori.

Citation: Journal of Hydrometeorology 15, 3; 10.1175/JHM-D-13-062.1

In a country such as Ethiopia, where rain-fed agriculture sustains 85% of the population, reliable seasonal forecasts have strong potential benefits (Diro et al. 2011a). However, both analysis and prediction of rainfall in this region is a challenge because of the complexity of the rainfall regime (Segele et al. 2009a) and the resultant myriad of factors governing its interannual variability. Much of Ethiopia has a summer rainfall regime, with maximum rainfall during the period June–September. The south and east, along with most of Somalia and Kenya, have two rainy seasons falling in the two transition seasons. However, most of the Horn of Africa receives some rainfall in each of the three seasons, and each season can contribute to the notable drought and flood situations (Nicholson 2013). The rainfall pattern is further complicated by topography and by the presence of Lake Victoria, which produces its own circulation system and rainfall regime (Ba and Nicholson 1998).

Both statistical and dynamical models have been applied to seasonal forecasting in this region, with the statistical models providing greater skill (Diro et al. 2011b; van Oldenborgh et al. 2005). The statistical models exploit teleconnections between rainfall and other variables. Most relate rainfall to surface parameters, generally sea surface temperatures (SSTs). A notable exception is the forecast model of Camberlin and Philippon (2002), which includes among the predictors u and υ winds and geopotential height. This current study differs from most previous ones in that 1) two geographical regions have been distinguished based on the rainfall regime rather than geographic considerations and 2) the three aforementioned rainy seasons are considered in both regions. In addition, an emphasis is placed on atmospheric circulation parameters, such as zonal winds at various levels. Several studies (e.g., Segele et al. 2009a,b; Hastenrath et al. 2004; Dezfuli and Nicholson 2013; Nicholson and Dezfuli 2013) have shown that these parameters can govern interannual variability in equatorial Africa, independent of their relationship to SSTs.

The geographic and climatic background of this region is considered in section 2, and previous forecast attempts are reviewed in section 3. Following a discussion of data and methodology in sections 4 and 5, forecast results are presented for each season in sections 6, 7, and 8. Section 9 presents the result of cross validation, and section 10 summarizes the results and conclusions.

2. Climatic background

Figure 1 depicts the rainfall regime over the Horn of Africa. Areas to the north have a summer maximum and areas to the south have the typical equatorial pattern of two rainy seasons during the transition seasons. In regions with this bimodal regime, the main rains (termed the “long rains”) occur during the boreal spring, generally from March to May. The second rainy season, the “short rains” of October and November, is generally briefer and less intense. However, it is the period of maximum rainfall in some areas of eastern Kenya.

For the region as a whole (Fig. 2), January–February is the driest period. In the equatorial rainfall region, June–September is also extremely dry. During March–May (MAM), significant rainfall occurs throughout the region.

Fig. 2.
Fig. 2.

Mean monthly rainfall (mm) over the study area.

Citation: Journal of Hydrometeorology 15, 3; 10.1175/JHM-D-13-062.1

The low-level winds (Fig. 3) are illustrated with data for 925 mb. This level is used instead of the surface because of the region’s terrain. Flow at 1000 and 850 mb (not shown) is very similar. During the MAM season, the low-level winds are southeasterly over most of the Horn of Africa. However, near the coast of East Africa, there is a progressive directional shift from northeast in March to southwest in May (not shown), the southwest flow being established as the easterlies over the subtropical Indian Ocean move toward the equator. From June to September, the mean low-level wind pattern is relatively steady, with southeast trades south of the equator becoming southwesterly upon transit into the Northern Hemisphere. The low-level Somali jet is well established during this season. The southwesterlies disappear in October, with a pattern of northeast and southeast trades being established in November, giving way to weak northeasterlies in December–February (not shown).

Fig. 3.
Fig. 3.

Mean vector winds by season for (top) 925, (middle) 700, and (bottom) 200 mb: (left) MAM, (middle) JAS, and (right) ON.

Citation: Journal of Hydrometeorology 15, 3; 10.1175/JHM-D-13-062.1

The flow is very different at higher levels (Fig. 3). In the midtroposphere (700 mb) the flow over the Horn of Africa is easterly or northeasterly in both the MAM and October–November (ON) seasons. It is very weak in the July–September (JAS) season, when westerlies prevail at 700 mb over the central and western equatorial Indian Ocean. In the upper troposphere (200 mb), easterlies prevail over the Indian Ocean throughout all three seasons, but they become very strong during JAS, when the tropical easterly jet (TEJ) is fully developed. In contrast, 200-mb flow is very weak over the Horn of Africa during the MAM and ON seasons.

Traditionally, the seasonal cycle is described as a result of the north–south excursion of the intertropical convergence zone (ITCZ). Its twice-yearly equatorial passage is assumed to be the cause of the bimodal equatorial rainfall regime. However, a look at the divergence at 925 mb suggests an alternative scenario. The October pattern is very typical (Fig. 4). The ITCZ, which extends across West Africa at roughly 15°N, lies over southern Sudan and the extreme northern tip of Ethiopia, but not over other sectors of the Horn of Africa. The narrow band of convergence zone over Kenya, Uganda, and Tanzania is a separate feature. The more western sector is associated with the Congo air boundary (Nicholson 1996), while that in the northeastern Horn is associated with the convergence of the trades over the western Indian Ocean.

Fig. 4.
Fig. 4.

Low-level (925 mb) divergence in October (10−6 s−1).

Citation: Journal of Hydrometeorology 15, 3; 10.1175/JHM-D-13-062.1

3. Prior studies of predictability of rainfall in the Horn of Africa

A number of studies have examined the predictability of rainfall in the Horn of Africa in each of the three seasons. Empirical methods, which rely upon statistical models of teleconnections among variables, are most commonly used. In general, the greatest predictability is for the summer season in the summer rainfall region and for the short rains of the boreal autumn in the equatorial rainfall region. Few of the studies are directly comparable because the time periods, season definition, predictors, geographical sectors, and approaches are quite varied. However, all demonstrate the potential predictability of seasonal rainfall in eastern Africa.

a. Boreal summer season

Termed the kiremt rains, the June–September season accounts for some 50%–80% of the rainfall over Ethiopia’s agricultural regions. Korecha and Barnston (2007) found that interannual variability during this season is governed primarily by ENSO, but that more local factors near Africa and in the Atlantic and Indian Oceans play a role as well. Using mainly multivariate statistical techniques, they obtained significant forecast skill, but with relatively short lead times limited by the spring predictability barrier of ENSO (e.g., Torrence and Webster 1998; Wajsowicz 2007). Their regression model used three variables to predict June–September rainfall: February–March SSTs in the South Atlantic and in the Niño-3.4 region plus May SST in the Niño-3.4 region. The model explained 59% of the overall rainfall variance and 41% of the variance in the cross-validation mode. Gissila et al. (2004) similarly utilized statistical forecasting for this season, with SSTs in the Pacific and Indian Oceans as predictors. They found that predictive skill could be improved by individually considering several regional sectors of the country. The degree of skill was also found to be regionally dependent.

Diro et al. (2011a,b) also used surface parameters to predict Ethiopian summer rains. In agreement with Gissila et al. (2004), they found that the best predictors vary by region, indicating that statistical forecast models need to take into account local variations in teleconnections. They found relationships between rainfall in six homogeneous regions of the country and SSTs in the equatorial Pacific, midlatitude regions of the northwestern Pacific, and the Gulf of Guinea. An important part of their study was to provide a justification for the use of predictor regions (such as the northwestern Pacific) for which no clear physical relationship to Ethiopia is apparent. Such predictors, which we also found, can be precursors to the development in later seasons of predictors with a direct physical connection.

Diro et al. (2012) further evaluated the potential of dynamical prediction of Ethiopian rainfall by using the Regional Climate Model, version 3 (RegCM3), to downscale European Centre for Medium-Range Weather Forecasts (ECMWF) seasonal ensemble forecasts. They found that the skill of probabilistic forecasts was greater with ECMWF on a gridpoint-by-gridpoint comparison than for their six homogeneous zones. However, RegCM3 had higher skill at the countrywide scale.

Block and Rajagopalan (2007), like the aforementioned studies, used surface variables as predictors in their model for rainfall in the upper Blue Nile basin (northwestern Ethiopia). In contrast, Segele et al. (2009a,b) demonstrated that large-scale regional atmospheric circulation patterns provide further forecast skill for Ethiopian summer rainfall. Considering the JAS season, they found that the strongest links are to various components of the African–Asian monsoon system, such as the Azores and Saharan highs, Mascarene high over the Indian Ocean, the TEJ, and the monsoon trough. Anomalously high rainfall was shown to be associated with enhanced westerlies across western and central Africa, an anomalously strong northeast-directed pressure gradient between the Gulf of Guinea and the Arabian Peninsula, deep moist air extending into the midtroposphere, large water vapor transport convergence across much of Ethiopia, a strong Somali low-level jet, and a strong TEJ. For the TEJ, the highest correlation was apparent at the 100-mb level.

b. The short rains of the boreal autumn

The short rains, although the secondary season in most of eastern equatorial Africa, provide the largest contribution to interannual variability. They also have one of the strongest associations ever demonstrated to global circulation: the correlation between East African rainfall during this season and the surface westerlies over the equatorial Indian Ocean is −0.85 (Hastenrath et al. 1993). This suggests a significant degree of predictability, assuming a fair degree of persistence of circulation parameters. Statistical forecast models for this season were developed by Philippon et al. (2002), Mutai et al. (1998), Ntale et al. (2003), Mwale and Gan (2005), and Hastenrath et al. (2004). Batte and Deque (2011) also examined the predictability of this season, but instead utilized numerical models. They evaluated both deterministic (single model) predictions and probabilistic (multimodel) skill scores.

The main months of the short rains are October and November. It should be noted that few of the studies were confined to these months. While it is well known that rainfall variability is highly coherent within the ON period, it is not clear whether or not the variability is coherent within the longer seasons (September–December or October–December) used by several studies. Camberlin and Philippon (2002) found that the coherence is limited to ON, but Hastenrath et al. (2004) found that for the coastal region the correlation between the ON season and the September–December season is 0.97. However, Hastenrath et al. (1993) show much greater skill in predicting October and November rainfall in this region than December rainfall.

Philippon et al. (2002) used a multiple linear regression model to predict the October–December rainfall in a large sector of East Africa that included inland Kenya, northern Tanzania, plus most of Rwanda, Burundi, and Uganda. Based on September predictors identified from correlations for the 1968–97 period, their model explained 64% of the interannual variance. The predictors included a monsoon index involving the northeast and southwest wind components at 200 and 850 mb, respectively; meridional wind at 200 mb over the southeastern tip of Africa; and an index of circulation over the western Indian Ocean.

Mutai et al. (1998) found that the JAS global SST pattern is strongly correlated with October–December seasonal rainfall aggregated for a large sector of East Africa extending from 5°N, in Kenya, southward to Malawi at 15°S. They developed a multiple linear regression forecast model based on three rotated EOFs for SSTs in the northwestern Pacific, the eastern equatorial Pacific (the ENSO signal), and the South Atlantic. The model showed significant forecast skill, with a correlation between predicted and observed of 0.69 for the period 1945 to 1988 for rainfall averaged over the entire region. The strongest predictor was an SST EOF with maximum variance in the northwest Pacific.

Ntale et al. (2003) used canonical correlation analysis to predict standardized seasonal rainfall totals for September–November at 3-month lead time. Predictors included sea level pressure (SLP) and SST anomaly fields in the Indian and Atlantic Oceans. The strongest association was with SSTs off the Somali and Benguela coasts. Mwale and Gan (2005) continued the work, comparing several methods of predicting standardized seasonal precipitation at 21 stations within a homogeneous region that comprises most of East Africa. Skill was higher with a nonlinear model known as artificial neural network than with the more standard linear canonical correlation model. In the latter case, the percent variance explained at individual stations for the 11 seasons 1987 and 1997 ranged from 49% to 81%, with root-mean-square error (RMSE) of 0.4–0.75 standardized units. With linear correlation the model explained 6% to 32%, with RMSE of 0.4–1.2.

Hastenrath et al. (2004) conducted several prediction experiments using a linear forecast model and a variable number of predictors, including two experiments with the Southern Oscillation index as the only predictor. The best predictors were zonal temperature and pressure gradients across the equatorial Indian Ocean. A cross validation for 1958–96 based on six predictors, produced a correlation between predicted and observed rainfall of 0.45. However, when the model was tested using separate training and validation periods, correlation in the validation period was much lower. It also appeared that the correlation with individual predictors changed markedly over time.

c. The long rains of the boreal spring

The boreal spring is the main rainy season in most of Kenya, Uganda, Somalia, and northern Tanzania. This season is termed the masika in Kenya–Uganda and gu in Somalia. The northern protrusion of these rains into Ethiopia is locally termed the belg (or small rains) season.

The most extensive study of the predictability of the boreal spring rains is that of Camberlin and Philippon (2002). As in the current study, they distinguished two geographical regions, separately considering Ethiopia and Kenya–Uganda, but predictability was tested only for the latter region. Four February indices, involving several time scales and both atmospheric and oceanic parameters, served as predictors in linear multiple regression and linear discriminant analysis models. The predictors were SST in Niño-1.2, zonal wind over the Congo basin at 1000 mb, geopotential height of the 500-mb surface over the Near East, and the east–west moist static energy gradient between the East African highlands and the Sahel. The models were applied for the period 1951–97 and were evaluated using cross validation. For the multiple regression model, the correlation between the predicted and observed MAM rainfall for the Kenya–Uganda section was 0.66 in the cross-validation mode. The discriminant analysis model correctly classified the seasonal anomalies 70% of the time.

Diro et al. (2008) and Ntale et al. (2003) also used empirical methods to predict rainfall in the boreal spring over East Africa. Both studies focused on Ethiopia. The latter study applied canonical correlation analysis to predict standardized MAM rainfall totals at a 3-month lead time, using SLP and SST anomaly fields of the Indian Ocean adjacent to East Africa and in the Gulf of Guinea in the Atlantic. Camberlin and Philippon (2002) similarly found strong local influence (the Red and Arabian Seas) on MAM rainfall in Ethiopia.

Because of the spatial variation across Ethiopia in both interannual variability and annual cycle, Diro et al. (2008) identified five homogeneous rainfall zones within Ethiopia and produced separate forecast models for each zone. Both multiple linear regression and linear discriminant analysis were applied to four sets of predictors. The study found that the method of selecting predictors had little impact on forecast skill. It was also shown that the models had the most skill in the southern and eastern parts of Ethiopia and that the extreme years were more reliably forecast than the average years.

4. Data and methodology

The current study is confined to the period 1950–2005. Rainfall data are taken from the author’s gauge dataset (e.g., Nicholson 1986, 2000). The stations included are shown in Fig. 5.

Fig. 5.
Fig. 5.

Stations used in the rainfall analysis with the black dots identifying a summer rainfall region and blue, an equatorial one.

Citation: Journal of Hydrometeorology 15, 3; 10.1175/JHM-D-13-062.1

The remaining data include SSTs and various atmospheric circulation parameters. SST data are from the National Oceanic and Atmospheric Administration (NOAA) Extended Reconstructed SST, version 3, product, which is available at a 2° × 2° resolution (Smith et al. 2008). The circulation data (surface pressure, zonal and meridional wind at various levels, and vertical motion omega) are taken from the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) Reanalysis (Kalnay et al. 1996; Kistler et al. 2001). The spatial resolution is 2.5°. Anomalies are calculated with respect to the 1998–2011 mean. A comparison was made with the 40-yr ECMWF Re-Analysis (ERA-40) and the circulation parameters showed reasonable agreement in the two datasets. NCEP is used for consistency with the author’s previous publications. The exception is divergence, which is calculated from Interim ECMWF Re-Analysis (ERA-Interim) data (Dee et al. 2011).

In a companion study (Nicholson 2013), two geographical regions were defined, based on the seasonal cycle of rainfall (Fig. 1). The areas where the rainfall maximum falls within the months of June–September were designated collectively as the summer rainfall region. Those with a maximum falling within the MAM season or in the ON season are designated the equatorial rainfall region. That study was based on high-resolution Tropical Rainfall Measuring Mission (TRMM) 3B42 rainfall estimates, allowing for a sharp delineation of regional boundaries. Our major regional partitioning agrees well with the regions identified by Diro et al. (2008), Gissila et al. (2004), and Tsidu (2012).

Here gauge data are used, and for each season a single time series is calculated for each of the two regions, based on the stations that fall within regional boundaries. The rainy seasons examined are MAM (boreal spring), JAS (boreal summer), and ON (boreal autumn). These seasonal divisions were selected based on changes in the spatial distribution of rainfall and on atmospheric circulation patterns. However, they correspond to the classic definitions of the rainy seasons in the Horn of Africa (e.g., Camberlin and Philippon 2002). Although June is usually included as part of the summer rainy season, it is excluded in the current study because it is actually a transition month.

Regional rainfall is represented as a standardized departure from the long-term mean. It is obtained by first calculating at each station a seasonal mean and a seasonal standard departure, then dividing by the standard departure. Following Nicholson (1986), the seasonal average for the region and year is a simple arithmetic average of all available stations in the region. This is the approach used in many other previous studies (e.g., Nicholson 1993; Nicholson et al. 2000; Lamb and Peppler 1992; Lebel and Ali 2009; Philippon et al. 2002). It facilitates the use of records with missing data, variable record lengths, and diverse climatologies. The remaining variables were similarly standardized.

5. Seasonal forecasting

In this study, the approach to seasonal forecasting has three stages: the selection of predictors; the development of a statistical forecast model for each region and season using multiple linear regression; and evaluation of forecast skill. When feasible, forecasts were developed for two lead times and in all cases, a constant lead time was used for the predictors. Forecast models were developed for the MAM, JAS, and ON seasons.

The approach was to commence with simple linear correlation, obtaining correlations between seasonal rainfall and atmospheric or oceanic variables. The latter include SST, SLP, omega (Ω), and zonal and meridional winds at 925, 850, 700, and 200 mb. These variables were chosen because they have been shown in previous studies to correlate with interannual variability in numerous locations (e.g., Camberlin 1995; Hastenrath and Polzin 2005; Dezfuli and Nicholson 2013; Nicholson and Dezfuli 2013). The wind and vertical motion fields were selected because they bear relationships to circulation features, such as jet streams and the Walker circulation, that appear to modulate interannual variability over equatorial Africa.

Maps were produced showing correlations at every global grid point between 45°N and 45°S for the independent variable in question and for various lead times. These allowed us to define feasible lead times. Then, coherent areas of highly significant correlation were selected as candidate predictors. These were limited to areas where the correlation exceeded the 0.001 significance level (~0.33). In most cases, the correlation between candidate predictor and observed rainfall exceeded 0.40–0.50 for at least one grid point. The average value of the variable within the area selected was used as input to the regression. Rainfall was also correlated with ENSO, using the Niño-3.4 SST index and with the Indian Ocean dipole (IOD) mode index based on the Hadley Center SST dataset (http://www.jamstec.go.jp/frcgc/research/d1/iod/iod_home.html.en).

The selection of geographical areas for candidate predictors was based not only on the degree of correlation but also on the principles that variables used in the regression should generally be independent and that a relatively small number of variables should be used, in order to avoid problems with colinearity and overfitting. After an initial set of predictor candidates was selected, the sensitivity of the regression to each predictor candidate was tested, as was the correlation among predictors. Those to which the model showed little sensitivity were then eliminated, the goals being to reduce the number of predictors utilized and to ensure the use of robust predictors. Candidates were also removed if they were highly correlated with other predictors, that is, when the common variance exceeded 20%.

After the final set of predictors was identified, the regression models were validated using cross correlation (Elsner and Schmertmann 1994; Michaelsen 1987). This approach has been typically used in other statistical forecasting studies for East Africa (e.g., Thiaw et al. 1999; Philippon et al. 2002; Diro et al. 2008, 2011b). It is optimally used on small datasets, when the analysis period is too short to produce reliable training and verification periods. The essence of the method is that a regression model is developed for each year, based on all other years in the dataset. Thus, the prediction is independent of the predictand.

The predictor variables for each case are indicated in Tables 1a, 2a, and 3a, along with the correlation between each variable and the predictand (rainfall in the indicated region/season). The equations defining the regression model are indicated in Tables 1b, 2b, and 3b. Table 4 gives the results of the regression models and the cross validations.

Table 1.

Predictors for MAM rainfall, geographical sectors corresponding to them, correlations with observed rainfall, and symbols used in the regression models prescribed by Eqs. (1)(3).

Table 1.
Table 2.

Predictors for JAS rainfall, geographical sectors corresponding to them, correlations with observed rainfall, and symbols used in the regression models prescribed by Eqs. (4) and (5).

Table 2.
Table 3.

Predictors for ON rainfall, geographical sectors corresponding to them, correlations with observed rainfall, and symbols used in the regression models prescribed by Eqs. (6)(9).

Table 3.
Table 4.

Summary of results of the regression models for prediction and cross validation.

Table 4.

6. Results for the boreal spring season (MAM)

For the MAM season, significant precursors could not be identified more than 2 months in advance. For the equatorial rainfall region, two regression models were developed, one based on February input data and the other based on January input data. For the summer rainfall region, there were no significant correlations prior to February, so that only one regression model was developed. Also, there were no geographical areas with correlations reaching the 1% level, so that the initial set of predictors utilized variables correlating with rainfall at the 5% level (roughly 0.26) or better.

The predictor variables are indicated in Table 1 and the location is shown in Fig. 6. The regression models for the three cases are indicated below:
e1
e2
e3
The meanings of the above symbols are given in Table 1.
Fig. 6.
Fig. 6.

Location of predictors (colors, see Table 1) used for the MAM season: (top) equatorial rainfall region and (bottom) summer rainfall region.

Citation: Journal of Hydrometeorology 15, 3; 10.1175/JHM-D-13-062.1

The regression models using February input data (Table 1a) produce a correlation between predicted and observed rainfall of 0.76 for the equatorial region and 0.63 for the summer rainfall region. Figure 7 shows the time series of observed versus predicted rainfall for the two regions. In the cross-validation mode, the correlations fell to 0.63 and 0.49, respectively. These results are comparable to those of Camberlin and Philippon (2002), who found a correlation between February predictors and Kenya–Uganda rainfall of 0.66 in the validation mode. However, their results were based on a shorter time period (1968 to 1997, compared to 1950 to 2005 in this study). The model for the equatorial rainfall region based on January input data (Table 1b) gives a correlation of 0.74 between predicted and observed rainfall, roughly equal to the correlation based on February input data (Fig. 7). In all three cases the RMSE ranges between 0.30 and 0.32 standardized units. Zonal and meridional winds provide greater predictive skill than SSTs or SLP.

Fig. 7.
Fig. 7.

Predicted (red) vs observed (black) MAM rainfall for the equatorial region based on (top) February and (middle) January predictors; and (bottom) for the summer rainfall region based on February predictors. Correlation r between the predicted and observed is indicated in each panel.

Citation: Journal of Hydrometeorology 15, 3; 10.1175/JHM-D-13-062.1

7. Results for the boreal summer season (JAS)

For the summer rainy season, forecast skill was not evident beyond a 2-month lead time. The predictor variables are indicated in Table 2, their location in Fig. 8, and the time series of predicted versus observed rainfall in Fig. 9. The regression models based on May input data are
e4
e5
The meanings of the variables are given in Table 2.
Fig. 8.
Fig. 8.

Location of predictors (colors, see Table 2) used for the JAS season for the equatorial and summer rainfall regions.

Citation: Journal of Hydrometeorology 15, 3; 10.1175/JHM-D-13-062.1

Fig. 9.
Fig. 9.

Predicted (red) vs observed (black) JAS rainfall based on May predictors: (top) equatorial and (bottom) summer regions. Correlation r between the predicted and observed is indicated in each panel.

Citation: Journal of Hydrometeorology 15, 3; 10.1175/JHM-D-13-062.1

The regression models produce a correlation between predicted and observed rainfall of 0.54 for the equatorial region and 0.81 for the summer rainfall region (Table 4). The low correlation for the equatorial region is not surprising, in that summer rainfall in the equatorial region is just an occasional spillover of the summer rainfall regime (Camberlin and Philippon 2002). RMSE is 0.32 and 0.25 standardized units, respectively, for the two regions.

For the summer rainfall region, SST, SLP, and zonal wind during May all provide high forecast skill during the boreal summer season. The correlations with zonal wind in the western equatorial Indian Ocean and SLP in the central equatorial Indian Ocean are −0.61 and −0.57, respectively, for the 56-yr period (Table 2a). The SST gradient in the Pacific is also well correlated with summer rainfall, with r = −0.71. Korecha and Barnston (2007) found ENSO to be the major predictor for this region, finding a correlation of −0.76 between Ethiopian summer rainfall and Niño-3.4 for the time period 1970–2004. In the current study, a high correlation with ENSO (−0.70) was also found, but predictability was enhanced using other indicators that are closely tied to ENSO. While ENSO explains 49% of the variance of JAS rainfall in the summer rainfall region, the regressive model in Eq. (5) explains 62% of the rainfall variance.

For the equatorial rainfall region [Eq. (4)], only four predictors could be identified (Table 2). These include two areas of SSTs, an area of SLP, and one zonal wind parameter. Compared to the summer rainfall region during this season, correlations between predictors and rainfall are low, ranging from −0.34 to +0.42. The SLP and SST predictors are all in the northern equatorial Pacific, while the zonal wind predictor (700-mb zonal wind) is located over the Mediterranean.

8. Results for the ON season

The greatest predictability is for the ON season. For both the equatorial rainfall region and the summer rainfall region there is significant skill with a 5-month lead time. It is roughly equivalent to, but somewhat higher than, the forecast skill with a 2-month lead time. Table 3 shows the predictors in the regression model and the correlation of each with ON rainfall. Figure 10 shows the location of the predictors.

Fig. 10.
Fig. 10.

Location of predictors (Table 3) used for the ON season: (top) equatorial rainfall and (bottom) summer rainfall regions. Red lines indicate that the predictor is the difference between the variable in the two liked boxes.

Citation: Journal of Hydrometeorology 15, 3; 10.1175/JHM-D-13-062.1

The model equations for the four regression models are indicated below:
e6
e7
e8
e9
The meanings of the variables are given in Table 3.

For the equatorial rainfall region, the overall correlation between rainfall predicted by the regression model and observed rainfall (Fig. 11) is 0.78 for a 2-month lead time and 0.80 for a 5-month lead time, respectively (Tables 3a,b). The RMSE is 0.43 and 0.40 standardized units, respectively. For the summer rainfall region (Fig. 11), the correlation is 0.77 and 0.72 for the 2- and 5-month lead times, respectively (Tables 3c,d). RMSE is 0.31 and 0.34.

Fig. 11.
Fig. 11.

Predicted (red) vs observed (black) ON rainfall for the (top two panels) equatorial and (bottom two panels) summer rainfall regions with August predictors in one case (first and third panels) and May predictors in a second case (second and fourth panels). Correlation r between the predicted and observed is indicated in each panel.

Citation: Journal of Hydrometeorology 15, 3; 10.1175/JHM-D-13-062.1

For prediction based on August variables (2-month lead time) the highest correlation is with the SST difference across the equatorial Indian Ocean, a measure of the IOD (r = −0.64). With this exception, correlations with SSTs per se and with SLP are notably weaker than with the wind field. For the equatorial region, the positive correlation with the difference in 925-mb zonal winds between the western equatorial Pacific and the eastern equatorial Indian Ocean (r = −0.59), together with the negative correlation with the difference in 200-mb zonal winds over the western and eastern Pacific Ocean (r = −0.50) suggests a link to a weakening of the Pacific Walker cell. This is consistent with the known positive relationship between El Niño events and the short rains of East Africa.

For prediction based on May variables (5-month lead time), the correlations between rainfall and predictors (Tables 3b,d) are notably lower than the correlations between rainfall and August predictors. Also, the correlation with surface variables (SSTs and SLP) is on the same order as that with higher-level atmospheric parameters. The regression model for the equatorial region includes only surface parameters. For the equatorial region, the highest correlation is with surface pressure over a large sector of the tropical North Atlantic.

9. Cross validation

Table 4 shows the results of cross validation for each of the regression models produced. Notably, in most cases the RMSE is only marginally higher in the cross validations than for the regression models. It was also noted that the regression coefficients remained stable in the cross validation, suggesting real physical relationships between the predictors and predictands, as opposed to random statistical associations. However, two of the models explained little of the variance and the variance explained dropped markedly in the cross validation.

The most robust regression models appear to be those for the ON season. For the equatorial rainfall region, the correlation between model forecasts based on May predictors and observed rainfall is 0.80, compared to a correlation of 0.71 using cross validation. For the summer rainfall region, the respective correlations for model and cross validation are 0.77 versus 0.68 when August predictors are used. Results are similar for the summer rainfall region (Fig. 12) during the JAS season. The correlation between the model prediction and observed rainfall is 0.79 versus a correlation of 0.75 in the cross validation.

Fig. 12.
Fig. 12.

Example of cross validation for (top) a case with high correlation between predicted (red) and observed (black) (summer rainfall region, JAS rainfall, and May predictors) and (bottom) a case with a low correlation between predicted and observed (equatorial rainfall region, JAS rainfall, and May predictors). Correlation r between the predicted and observed is indicated in each panel.

Citation: Journal of Hydrometeorology 15, 3; 10.1175/JHM-D-13-062.1

The cases where the cross validation suggests that the forecast model is not robust are those where the model explains the lowest percent variance: the equatorial rainfall region in the JAS season and the summer rainfall region in the MAM season. Figure 12 shows the cross-validation results for the former case. Clearly, the model is inadequate in this case.

10. Summary and discussion

The results of the regression models are summarized in Table 4. The number of predictors ranged from 3 to 7. The models based on 2- and 5-month lead time showed considerable skill in predicting rainfall for the ON season for both the summer and equatorial rainfall regions and in predicting JAS rainfall in the summer rainfall region. The correlations between predicted and observed rainfall ranged from 0.54 to 0.80, with RMSE on the order of 0.25–0.35 for the summer rainfall region and 0.3–4 for the equatorial region. For ON the predictive skill was roughly the same for 2- and 5-month lead times. Predictive skill for the JAS season was notably lower for the equatorial region. However, this is a period that is relatively dry throughout most of that region.

Notably, potential predictors could not be identified for either season prior to May. This result is consistent with the spring “prediction barrier” noted by Torrence and Webster (1998), Wajsowicz (2007), and others. This phenomenon is related mainly to the state of the tropical Pacific in the boreal spring (Webster and Yang 1992; Ndiaye et al. 2009), a time of transition between climate states and weak SST gradients. Consequently, the signal-to-noise ratio is at a minimum and the system readily responds to perturbations (Torrence and Webster 1998). Whether these perturbations persist is determined in late spring or early summer, a time when ENSO events rapidly evolve. Similar constraints may hold for the Indian Ocean and IOD (Webster et al. 1999). Other factors contributing to the predictability barrier over East Africa include the rapid change over circulation within the boreal spring season and marked changes in teleconnections between March–April and May (Camberlin and Philippon 2002). Korecha and Barnston (2007) likewise noted that this barrier severely limits the possibility of East African seasonal forecasting 2 or more months in advance.

The regression models showed less success for the MAM season. For the equatorial region, predicted and observed rainfall correlated well with 1- and 2-month lead times, but not with longer lead times. For the summer rainfall region, which receives considerable rainfall in the MAM season as well, predictive skill was not evident beyond 1 month in advance and was still low, compared to the other cases we evaluated.

Camberlin and Philippon (2002) also found relatively poor predictability of the MAM season and suggested that this is a result of the lack of intermonth coherency within this season. They noted that the dynamic factors operative in May are markedly different from those in March and April and suggested that May be evaluated separately. Moreover, the links to potential predictors are quite different for each of the 3 months.

Two notable results of the current study are that the predictors are considerably different for different lead times and that upper-level atmospheric circulation provides greater predictability than surface conditions. SSTs and SLP were most useful for the summer rainfall region and for prediction of ON rainfall with a long lead time.

The most useful predictor was zonal wind, particularly for the summer rainfall region. Based on this result and the original correlation maps, it appears that changes in Walker-like zonal circulation cells are critically important. Confirmation of the role of zonal cells comes from high correlations of the opposition sign with low-level and upper-level zonal winds at the same location. Dezfuli and Nicholson (2013) and Nicholson and Dezfuli (2013) reached a similar conclusion for western equatorial Africa.

The Azores (North Atlantic) high also appears to be an important precursor for ON rainfall in both the summer and equatorial rainfall regions. Interestingly, changes in that high appear to be precursors to ENSO as well (Nicholson 1997). Meridional winds were most important for predicting MAM rainfall for the equatorial rainfall region.

Both the IOD and ENSO appeared to be important precursors of rainfall in the ON season. Surprisingly, the IOD and ENSO indices were each selected as a predictor in only one of the nine cases. This does not indicate a lack of a relationship. Instead, variables closely linked to either the IOD or ENSO, such as SLP or SST gradients or zonal winds, provided more predictive skill than either index.

Without further detailed study, it is only possible to speculate on why atmospheric variables provide greater predictability than surface variables. However, it is to be noted that, except for coastal stations, SSTs represent remote forcing. The patterns of tropospheric convergence–divergence and vertical motion provide the direct forcing and are closely tied into the dominant stability mechanisms in the tropics (e.g., combined baroclinic–barotropic, inertial instability). Circulation patterns are crucial to these instabilities. Further, while SSTs may force these atmospheric variables that modulate seasonal rainfall, they also respond to them. This would weaken any direct links between surface variables, especially SSTs. The relationships to SSTs are also less stable over time than the relationships to atmospheric variables. Finally, atmospheric variables incorporate important factors such as aerosols and vegetation that SSTs do not.

There are two open questions concerning the results of our statistical approach, which provides bulk predictions for a relatively large region. These are whether the model provides adequate predictions throughout the region and whether it is adequate for the recent period, in which there appears to have a significant shift in the precipitation regime over eastern Africa.

In terms of the first question, the two analysis sectors are known to be relatively homogeneous with respect to interannual variability. However, several authors (e.g., Friederichs and Paeth 2006; Diro et al. 2008, 2012) have suggested that performance can be improved by examining subregions within East Africa. We tested this by isolating a sector of the equatorial region that Camberlin and Planchon (1997) identify as having a significantly different rainfall regime than the rest of Kenya. This is the area of central highlands where topographic effects are pronounced. The correlation between a time series based on nine stations in this region and the whole equatorial rainfall regime is 0.84 for the MAM season and 0.91 for the ON season. Figure 13 shows the results of the regression model applied to predicting rainfall in the highlands subregion for the two seasons. For the ON season, the correlation between observed and predicted is 0.76, compared to 0.78 for the whole equatorial rainfall region. For the MAM season the model performance is somewhat lower than for the whole region, 0.63 compared to 0.74 for the latter. Part of the difference can be accounted for by the greater variance of the time series, a result of the much smaller number of stations (9 vs 101).

Fig. 13.
Fig. 13.

Predicted (red) vs observed (black) rainfall for a subregion in the Kenyan highlands. (top) ON rainfall based on August predictors. (bottom) MAM rainfall based on January predictors. Correlation r between the predicted and observed is indicated in each panel.

Citation: Journal of Hydrometeorology 15, 3; 10.1175/JHM-D-13-062.1

The second question arises because several authors have noted recent changes in the precipitation regime over East Africa. Williams et al. (2012) show a shift in the relationship of Ethiopian summer rainfall to surface parameters such as SSTs and SLP around 1990. Lyon and DeWitt (2012) demonstrate an abrupt decline in the long rains that commenced around 1999. Nicholson (2013) has suggested that the character of droughts changed around 2005. A cursory answer to the question of model performance in recent years can be gleaned from our current results. For the equatorial regime there does not appear to be any systematic change in the predictability in either region after the suggested times of change (Figs. 7, 9, and 11). For the summer rainfall regime, the differences between predicted and observed appear to be larger during the last 10 years for the rainy seasons of the boreal spring and autumn, but not during the main rainy season of JAS.

Acknowledgments

The author was supported by NSF Grants AGS 1160750 and AGS 1158984. The programming contribution of Douglas Klotter is greatly acknowledged. Relevant discussions with Prof. Peter Webster are greatly appreciated.

REFERENCES

  • Ba, M. B., , and Nicholson S. E. , 1998: Analysis of convective activity and its relationship to the rainfall over the Rift Valley lakes of East Africa during 1983–90 using the Meteosat infrared channel. J. Appl. Meteor., 37, 12501264, doi:10.1175/1520-0450(1998)037<1250:AOCAAI>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Batte, L., , and Deque M. , 2011: Seasonal predictions of precipitation over Africa using coupled ocean-atmosphere general circulation models: Skill of the ENSEMBLES project multimodel ensemble forecasts. Tellus, 63A, 283299, doi:10.1111/j.1600-0870.2010.00493.x.

    • Search Google Scholar
    • Export Citation
  • Block, P., , and Rajagopalan B. , 2007: Interannual variability and ensemble forecast of upper Blue Nile basin kiremt season precipitation. J. Hydrometeor., 8, 327343, doi:10.1175/JHM580.1.

    • Search Google Scholar
    • Export Citation
  • Camberlin, P., 1995: June–September rainfall in north-eastern Africa and atmospheric signals over the tropics: A zonal perspective. Int. J. Climatol., 15, 773783, doi:10.1002/joc.3370150705.

    • Search Google Scholar
    • Export Citation
  • Camberlin, P., , and Planchon O. , 1997: Coastal precipitation regimes in Kenya. Geogr. Ann., 79A, 109119, doi:10.1111/1468-0459.00010.

  • Camberlin, P., , and Philippon N. , 2002: The East African March–May rainy season: Associated atmospheric dynamics and predictability over the 1968–97 period. J. Climate, 15, 10021019, doi:10.1175/1520-0442(2002)015<1002:TEAMMR>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Dee, D. P., and et al. , 2011: The ERA-Interim reanalysis: Configuration and performance of the data assimilation system. Quart. J. Roy. Meteor. Soc., 137, 553597, doi:10.1002/qj.828.

    • Search Google Scholar
    • Export Citation
  • Dezfuli, A. K., , and Nicholson S. E. , 2013: The relationship of interannual variability in western equatorial Africa to the tropical oceans and atmospheric circulation. Part II: The boreal autumn. J. Climate, 26, 6684, doi:10.1175/JCLI-D-11-00686.1.

    • Search Google Scholar
    • Export Citation
  • Diro, G. T., , Grimes D. I. F. , , and Black E. , 2008: Seasonal forecasting of Ethiopian spring rains. Meteor. Appl., 15, 7383, doi:10.1002/met.63.

    • Search Google Scholar
    • Export Citation
  • Diro, G. T., , Grimes D. I. F. , , and Black E. , 2011a: Teleconnections between Ethiopian summer rainfall and sea surface temperature: Part I—Observation and modelling. Climate Dyn., 37, 103119, doi:10.1007/s00382-010-0837-8.

    • Search Google Scholar
    • Export Citation
  • Diro, G. T., , Grimes D. I. F. , , and Black E. , 2011b: Teleconnections between Ethiopian summer rainfall and sea surface temperature: Part II: Seasonal forecasting. Climate Dyn., 37, 121131, doi:10.1007/s00382-010-0896-x.

    • Search Google Scholar
    • Export Citation
  • Diro, G. T., , Tompkins A. M. , , and Bi X. , 2012: Dynamical downscaling of ECMWF Ensemble seasonal forecasts over East Africa with RegCM3. J. Geophys. Res.,117, D16103, doi:10.1029/2011JD016997.

  • Elsner, J. B., , and Schmertmann C. P. , 1994: Assessing forecast skill through cross-validation. Wea. Forecasting, 9, 619624, doi:10.1175/1520-0434(1994)009<0619:AFSTCV>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Friederichs, P., , and Paeth H. , 2006: Seasonal prediction of African prediction with ECHAM4-T42 ensemble simulations using a multivariate MOS prediction scheme. Climate Dyn., 27, 761786, doi:10.1007/s00382-006-0154-4.

    • Search Google Scholar
    • Export Citation
  • Gissila, T., , Black E. , , Grimes D. , , and Slingo J. , 2004: Seasonal forecasting of the Ethiopian summer rains. Int. J. Climatol., 24, 13451358, doi:10.1002/joc.1078.

    • Search Google Scholar
    • Export Citation
  • Hastenrath, S., , and Polzin D. , 2005: Mechanisms of climate anomalies in the equatorial Indian Ocean. J. Geophys. Res.,110, D0811, doi:10.1029/2004JD004981.

    • Search Google Scholar
    • Export Citation
  • Hastenrath, S., , Nicklis A. , , and Greischer L. , 1993: Atmospheric-hydrospheric mechanisms of climate anomalies in the western equatorial Indian Ocean. J. Geophys. Res., 98, 20 219–20 235, doi:10.1029/93JC02330.

    • Search Google Scholar
    • Export Citation
  • Hastenrath, S., , Polzin D. , , and Camberlin P. , 2004: Exploring the predictability of the ‘short rains’ at the coast of East Africa. Int. J. Climatol., 24, 13331343, doi:10.1002/joc.1070.

    • Search Google Scholar
    • Export Citation
  • Kalnay, E., and et al. , 1996: The NCEP/NCAR 40-Year Reanalysis Project. Bull. Amer. Meteor. Soc., 77, 437471, doi:10.1175/1520-0477(1996)077<0437:TNYRP>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Kistler, R., and et al. , 2001: The NCEP–NCAR 50-Year Reanalysis: Monthly means CD-ROM and documentation. Bull. Amer. Meteor. Soc., 82, 247267, doi:10.1175/1520-0477(2001)082<0247:TNNYRM>2.3.CO;2.

    • Search Google Scholar
    • Export Citation
  • Korecha, D., , and Barnston A. G. , 2007: Predictability of June–September rainfall in Ethiopia. Mon. Wea. Rev., 135, 628650, doi:10.1175/MWR3304.1.

    • Search Google Scholar
    • Export Citation
  • Lamb, P. J., , and Peppler R. A. , 1992: Further case studies of tropical Atlantic surface atmospheric and oceanic patterns associated with sub-Saharan drought. J. Climate, 5, 476488, doi:10.1175/1520-0442(1992)005<0476:FCSOTA>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Lebel, T., , and Ali A. , 2009: Recent trends in the Central and western Sahel rainfall regime (1990–2007). J. Hydrol., 375, 5264, doi:10.1016/j.jhydrol.2008.11.030.

    • Search Google Scholar
    • Export Citation
  • Lyon, B., , and DeWitt D. G. , 2012: A recent and abrupt decline in the East African long rains. Geophys. Res. Lett., 39, L02702, doi:10.1029/2011GL050337.

    • Search Google Scholar
    • Export Citation
  • Michaelsen, J., 1987: Cross-validation in statistical climate forecast models. J. Climate Appl. Meteor., 26, 15891600, doi:10.1175/1520-0450(1987)026<1589:CVISCF>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Mutai, C. C., , Ward M. N. , , and Colman A. W. , 1998: Towards the prediction of the East Africa short rains based on sea-surface temperature–atmosphere coupling. Int. J. Climatol., 18, 975997, doi:10.1002/(SICI)1097-0088(199807)18:9<975::AID-JOC259>3.0.CO;2-U.

    • Search Google Scholar
    • Export Citation
  • Mwale, D., , and Gan T. Y. , 2005: Wavelet analysis of variability, teleconnectivity, and predictability of the September–November East African rainfall. J. Appl. Meteor., 44, 256269, doi:10.1175/JAM2195.1.

    • Search Google Scholar
    • Export Citation
  • Ndiaye, O., , Goddard L. , , and Ward M. N. , 2009: Using regional wind fields to improve general circulation model forecasts of July–September Sahel rainfall. Int. J. Climatol., 29, 12621275, doi:10.1002/joc.1767.

    • Search Google Scholar
    • Export Citation
  • Nicholson, S. E., 1986: The spatial coherence of African rainfall anomalies: Interhemispheric teleconnections. J. Climate Appl. Meteor., 25, 13651381, doi:10.1175/1520-0450(1986)025<1365:TSCOAR>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Nicholson, S. E., 1993: An overview of African rainfall fluctuations of the last decade. J. Climate, 6, 14631466, doi:10.1175/1520-0442(1993)006<1463:AOOARF>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Nicholson, S. E., 1996: A review of climate dynamics and climate variability in eastern Africa. The Limnology, Climatology and Paleoclimatology of the East African Lakes, T. C. Johnson and E. Odada, Eds., Gordon and Breach, 25–56.

  • Nicholson, S. E., 1997: An analysis of the ENSO signal in the tropical Atlantic and western Indian Oceans. Int. J. Climatol., 17, 345375, doi:10.1002/(SICI)1097-0088(19970330)17:4<345::AID-JOC127>3.0.CO;2-3.

    • Search Google Scholar
    • Export Citation
  • Nicholson, S. E., 2000: The nature of rainfall variability over Africa on time scales of decades to millennia. Global Planet. Change, 26, 137158, doi:10.1016/S0921-8181(00)00040-0.

    • Search Google Scholar
    • Export Citation
  • Nicholson, S. E., 2013: A detailed look the recent drought situation in the Great Horn of Africa. J. Arid Environ., 103, 71–79, doi:10.1016/j.jaridenv.2013.12.003.

  • Nicholson, S. E., , and Dezfuli A. K. , 2013: The relationship of interannual variability in western equatorial Africa to the tropical oceans and atmospheric circulation. Part I: The boreal spring. J. Climate, 26, 4565, doi:10.1175/JCLI-D-11-00653.1.

    • Search Google Scholar
    • Export Citation
  • Nicholson, S. E., , Some B. , , and Kone B. , 2000: An analysis of recent rainfall conditions in West Africa including the rainy seasons of the 1997 El Niño and the 1998 La Niña years. J. Climate, 13, 26282640, doi:10.1175/1520-0442(2000)013<2628:AAORRC>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Ntale, H. K., , Gan T. Y. , , and Mwale D. , 2003: Prediction of East African seasonal rainfall using simplex canonical correlation analysis. J. Climate, 16, 21052112, doi:10.1175/1520-0442(2003)016<2105:POEASR>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Philippon, N., , Camberlin P. , , and Fauchereau N. , 2002: Empirical predictability study of October–December East African rainfall. Quart. J. Roy. Meteor. Soc., 128, 22392256, doi:10.1256/qj.01.190.

    • Search Google Scholar
    • Export Citation
  • Segele, Z. T., , Lamb P. J. , , and Leslie L. M. , 2009a: Seasonal-to-interannual variability of Ethiopia/Horn of Africa monsoon. Part I: Associations of wavelet-filtered large-scale atmospheric circulation and global sea surface temperature. J. Climate, 22, 33963421, doi:10.1175/2008JCLI2859.1.

    • Search Google Scholar
    • Export Citation
  • Segele, Z. T., , Lamb P. J. , , and Leslie L. M. , 2009b: Large-scale atmospheric circulation and global sea surface temperature associations with Horn of Africa June–September rainfall. Int. J. Climatol., 29, 10751100, doi:10.1002/joc.1751.

    • Search Google Scholar
    • Export Citation
  • Smith, T. M., , Reynolds R. W. , , Peterson T. C. , , and Lawrimore J. , 2008: Improvements to NOAA’s historical merged land–ocean surface temperature analysis (1880–2006). J. Climate, 21, 22832296, doi:10.1175/2007JCLI2100.1.

    • Search Google Scholar
    • Export Citation
  • Thiaw, W. M., , Barnston A. G. , , and Kumar V. , 1999: Predictions of African rainfall on the seasonal timescale. J. Geophys. Res., 104, 31 58931 597, doi:10.1029/1999JD900906.

    • Search Google Scholar
    • Export Citation
  • Torrence, C., , and Webster P. J. , 1998: The annual cycle of persistence in the El Niño/Southern Oscillation. Quart. J. Roy. Meteor. Soc., 124, 19852004, doi:10.1002/qj.49712455010.

    • Search Google Scholar
    • Export Citation
  • Tsidu, M., 2012: High-resolution monthly rainfall database for Ethiopia: Homogenization, reconstruction, and gridding. J. Climate, 25, 84228443, doi:10.1175/JCLI-D-12-00027.1.

    • Search Google Scholar
    • Export Citation
  • van Oldenborgh, G. J., , Balmaseda M. , , Ferranta L. , , Stockdale T. , , and Anderson D. , 2005: Did the ECMWF seasonal forecast model outperform statistical ENSO forecast models over the last 15 years? J. Climate, 18, 32403249, doi:10.1175/JCLI3420.1.

    • Search Google Scholar
    • Export Citation
  • Wajsowicz, R. C., 2007: Seasonal-to-interannual forecasting of tropical Indian Ocean sea surface temperature anomalies: Potential predictability and barriers. J. Climate, 20, 33203343, doi:10.1175/JCLI4162.1.

    • Search Google Scholar
    • Export Citation
  • Webster, P. J., , and Yang S. , 1992: Monsoon and ENSO: Selectively interactive systems. Quart. J. Roy. Meteor. Soc., 118, 877926, doi:10.1002/qj.49711850705.

    • Search Google Scholar
    • Export Citation
  • Webster, P. J., , Moore A. , , Loschnigg J. , , and Leban M. , 1999: Coupled ocean–atmosphere dynamics in the Indian Ocean during 1997–98. Nature, 401, 356360, doi:10.1038/43848.

    • Search Google Scholar
    • Export Citation
  • Williams, A. P., and et al. , 2012: Recent summer precipitation trends in the Greater Horn of Africa and the emerging role of Indian Ocean sea surface temperature. Climate Dyn., 39, 23072328, doi:10.1007/s00382-011-1222-y.

    • Search Google Scholar
    • Export Citation
Save