Severe and recurrent rainfall deficits across the African continent during the past two to three decades have been detrimental to the economy of the African nations. Thus, policy makers and funding agencies often face tough challenges to make relief plans. There clearly is a need for forecasts of short-term climate fluctuations, such as for seasonal total rainfall one or more seasons in advance. The African Desk, established at the Climate Prediction Center (CPC) of the NWS/NCEP, has been experimenting with African seasonal forecasting in collaboration with the CPC. While numerical approaches are being considered, work so far has focused more on statistical methods. Here we apply canonical correlation analysis (CCA) to produce an experimental forecast for rainfall anomalies in the region of eastern Africa (10S-10N, 25E-50E) for the Oct-Nov-Dec 1996 period. The boreal spring and fall months comprise the climatologically rainy season in eastern Africa, as this is when the highest temperatures and atmospheric moisture content (i.e. the ITCZ) occur at this tropical region along the immediate equator.

The CCA method is a multivariate regression that relates patterns in predictor fields to patterns in the predictand field. The prediction design used here is the same as that of the CCA used as one of the tools for operational climate prediction in the U.S. (Barnston 1994), based on earlier work of Barnett and Preisendorfer (1987). Four consecutive 3-month predictor periods are followed by a lead time and then a single 3-month predictand, or target, period. Forecast skill experiments have indicated that the global SST field serves best as a predictor. While additional fields such as upper air geopotential height, tropical low-level wind or outgoing longwave radiation might well enhance skill farther, data sets of these fields do not extend far enough into the past to satisfy the CCA's need for a long-term (e.g. at least 25-year) data record from which to identify the dominant relationships. The predictor and predictand data sets used here begin in 1955. For the 1996 eastern African rainfall prediction shown below, the predictor data are the global SST anomaly field over the four 3-month periods of Sep-Oct-Nov 1995, Dec-Jan-Feb 1995-96, Mar-Apr-May and Jun-Jul-Aug of 1996. Using data from 1955-94, relationships between the prior year's SST anomaly evolution and the target year's Oct-Nov-Dec eastern African rainfall anomaly patterns are linearly modeled by the CCA. The predictor SST data for the current forecast are then projected onto the preferred relationships derived from the past years, and a forecast for the 1996 target season is developed. Here the lead time is 1 month, as the latest predictor data used are those of August 1996, preceding the beginning of the target period by 1 month.

The predictor SST data were derived from a combination of the COADS data (Slutz et al. 1985) and more recent OI data (Smith et al. 1996). The predictand Sahel rainfall data come from the gridded global rainfall data set developed by M. Hulme (Hulme 1994), at 2.5 by 3.75 latitude-longitude resolution, resulting in 29 points in eastern Africa (Fig. 1). A rainfall data set consisting of individual stations has also been tested, with results shown in Thiao et al. (1996) and Barnston et al. (1996). The gridded data used here tend to show slightly higher skill in parts of Africa having sparser data, such as eastern Africa. This may be because the gridded data are developed using stations that have major gaps during some periods, while the station data set completely excludes such stations.

The diagnostic data produced by CCA indicate that expected skill is mainly poor in predicting Oct-Nov-Dec precipitation in eastern Africa at 1-month lead, with average region wide correlation skill of 0.13, and 0.40 or higher skill at a couple of locations. Highest skills tend to occur in southern Sudan. A cross-validation design is used in obtaining these skill estimates, where each year is held out of the developmental data set in turn, and then used as the forecast target.

The forecast presented below is expressed in terms of anomalies (i.e., departures) from climato-logical probabilities of the occurrence of three equiprobable categories: below normal, near normal and above normal rainfall (B, N, and A, respectively). The climatological probability of each of these categories is 1/3; that is, over a long time period, each category is observed one-third of the time. The forecast provides the estimated deviation from these long-term probabilities for the period being forecast. For example, if a 0.07 probability anomaly for below normal rainfall is forecast, then the probabilities of the below, near and above normal categories to occur would be 0.403, 0.333 and 0.263 instead of the climatological probabilities of 0.333, 0.333, 0.333. In some cases the probability of the near normal category may be elevated at the expense of the two outer categories. Expressing the forecasts in terms of probability anomalies enables uncertainty to be incorporated into the forecasts. It also assumes knowledge on the part of the users regarding the rainfall amount intervals defined by the bottom, middle and top thirds of the distribution. In a rough sense, these categories may often be equated to drought, normal, or excessive rainfall cases.

The probability forecast for eastern Africa for Oct-Nov-Dec 1996 is shown in Fig. 2. The numbers on the map are probability anomalies for the above (A) normal category (X100). Slightly positive departures from climatological probabilities in the direction of above average rainfall for Oct-Nov-Dec are indicated for most of the locations having sufficient combinations of forecast skill and forecast signals to result in any departure. The field significance of the collective skill over all 29 grid points, computed using Monte Carlo randomizations (indicating the probability that the skill level of this forecast map could have occurred by chance), is 0.047. This satisfies the 0.05 level generally used as the threshold for field significance.

Although 6 CCA predictor-predictand modes are used in this prediction, the leading one to two modes typically contribute to the majority of the skill, especially when skill is modest as it is here. Mode 1, which describes an interdecadal SST trend, does not govern this forecast heavily because the recent SST patterns have not projected strongly onto this mode. The spatial loading patterns of mode 2, however, suggest that an important determinant of the current forecast is the SST anomaly in the eastern tropical Pacific (including but not limited to the ENSO state) and also a North Atlantic versus South Atlantic SST anomaly difference. A warm eastern tropical Pacific SST tends to be associated with positive rainfall anomalies in eastern Africa, as is the case for positive anomalies in the North Atlantic and negative anomalies in the South Atlantic. In the most recent several months, while the ENSO state has been mildly negative, the northern versus southern Atlantic SST anomaly pattern has been conducive to enhanced eastern African rainfall. The latter situation has been a stronger influence than the ENSO state in the current rainfall forecast.

The probability anomalies shown in Fig. 2 are all quite weak, with none exceeding 0.05. While this is mainly because of the modest skill at most of the locations, it is secondarily due to the neutrality of the forecast signal for this particular prediction. That is, the anomalies in the predictor SSTs are not strongly indicative of a well-defined anomaly pattern in the African rainfall for the coming several months. This makes sense in view of the lack of a strong overall anomaly in the tropic-wide SST over the last several months. While the tropical Pacific SST has continued to be slightly cooler than average, the influence of this anomaly has been outweighed by the positive SST anomalies in the Indian ocean and tropical and south-subtropical Atlantic ocean.

Barnston, A.G., W. Thiao and V. Kumar, 1996: Long-lead forecasts of seasonal precipitation in Africa using CCA. Wea. Forecasting, 11, in press.

Hulme, M., 1994: Validation of large-scale precipitation fields in general circulation models. In Global Precipitation and Climate Change, M. Desbois and F. Desalmand, Ed., NATO ASI Series, Springer-Verlag, Berlin, 466 pp.

Slutz, R., S.J. Lubler, J.D. Hiscox, S.D. Woodruff, R.J. Jenne, D.H. Joseph, P.M. Steurer, and J.D. Elius, 1985: Comprehensive Ocean Atmosphere Data Set. NOAA, Boulder, CO, 268 pp. [Available from Climate Research Program, ERL, R/E/AR6, 325 Broadway, Boulder, CO 80303.]

Smith, T.M., R.W. Reynolds, and C.F. Ropelewski, 1994: Optimal averaging of seasonal sea surface temperatures and associated confidence intervals (1860-1989). J. Climate, 7, 949-964.

Thiao, W., A.G. Barnston and V. Kumar, 1996: Teleconnections and seasonal rainfall prediction in Africa. Proceedings of the 20th Annual Climate Diagnostics Workshop, Seattle, Washington, October 23-27, 1995, 413-416.

Fig. 1. (top) The 29 grid points used as the rainfall predictand for the eastern African CCA predictions.

Fig. 2. (bottom) The CCA-based rainfall probability anomaly forecasts for the eastern African region of Africa for Oct-Nov-Dec 1996. Values shown are deviations from climatological probabilities for the above normal category. For example, 3 implies (B,N,A) probabilities of 0.303, 0.333, 0.363, while -2 implies 0.353, 0.333, 0.313. See text for further details.