[Previous Article] [Next Article]
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 Southern Africa (10-35S, 10-50E) for the Jan-Feb-Mar 1996-97 period. The austral summer months comprise the climatologically rainy
season in Southern Africa, as this is when the highest temperatures and atmospheric moisture content (i.e. the ITCZ)
occur at most parts of this tropical region of the Southern Hemisphere.
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 1996-97 Southern Africa rainfall prediction shown below, the predictor data are the global SST
anomaly field over the four 3-month periods of Dec-Jan-Feb 1995-96, Mar-Apr-May, Jun-Jul-Aug and Sep-Oct-Nov
1996. Using data from 1955-95, relationships between the prior year's SST anomaly evolution and the target year's Jan-Feb-Mar Southern Africa 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 1996-97 austral summer developed. A lead time of 1 month is used--the time by which the latest predictor data (November)
precede the target period onset.
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. 1994). The predictand Southern Africa rainfall data come from the gridded global rainfall data set
developed by M. Hulme (Hulme 1994), at 2.5 by 3.75 resolution, resulting in 58 points in the Southern Africa. 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). While skill results are roughly similar between the two rainfall data sets for the Sahel because
of the sufficient station data density, the gridded data tend to show higher skill in parts of Africa having sparser data.
This may be because the gridded data are developed using stations that may 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 modest in predicting Jan-Feb-Mar
precipitation at 1-month lead, with average region wide correlation skill of 0.24, and 0.50 or higher at some locations.
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.
Although 6 CCA predictor-predictand modes are used in this prediction, the leading 1 to 3 modes typically
accounts for most of the skill, especially when skill is not high as in the present case. The spatial loading pattern of mode
1 suggests that the tropical pacific SST anomaly, representing the ENSO state, is an important determinant of the current
forecast. A cool tropical Pacific SST tends to be associated with positive rainfall anomalies in most of southern Africa.
In the most recent several months, the ENSO state has been mildly cool and conducive to enhanced to near normal
rainfall in much of Southern Africa.
The forecast presented below is expressed in terms of probability anomalies with respect to climatological
probabilities for three equiprobable categories: below, near, and above normal (B, N, and A, respectively). The
climatological probability of each category is 1/3. The forecast provides the estimated deviation from these probabilities
for Jan-Feb-Mar 1997. For example, a 0.08 probability anomaly for above normal rainfall implies probabilities of the
below, near, and above normal categories of 0.253, 0.333 and 0.413 compared to the climatological probabilities of
o.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 probabilistically enables uncertainty to be conveyed in the forecasts. It
assumes that the users have some idea of the rainfall amount internals 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 plentiful rainfall cases.
The probability forecast for southern Africa for Jan-Feb-Mar 1997 is shown in Fig. 1. The predictand grid points
are shown by the locations of the numbers on the map. These numbers are probability anomalies for the above (A), near
normal (N), and below (B) categories (X100). Slightly positive departures from climatological probabilities (i.e.,
in the direction of above average rainfall) are indicated for much of Zambia, while dryness is predicted for
northern Namibia and southern Tanzania. In parts of southern Africa only climatological probabilities are
given, mainly because the historical forecast skill is too low. At some grid points this may occur also partly
because the predictor signal for this year's forecast is of insufficient strength. The field significance of the
collective skill over all 58 grid points, computed using Monte Carlo randomizations (indicating the probability that the
skill level of this forecast map could have occurred by chance), is 0.000. This strongly satisfies the 0.05 level generally
used as the threshold for field significance.
The probability anomalies shown in Fig. 1 are fairly weak, with no anomaly of 0.10 or higher. This is also because
of the modest skill at most of the locations, and secondarily due to the weakness of the forecast signal for this particular
prediction. The anomalies in the predictor SSTs are only weakly indicative of enhanced southern African rainfall by
virtue of the mild La Nina condition which may become a nearly neutral condition toward the end of the Jan-Feb-Mar
period.
Barnett, T.P. and R. Preisendorfer, 1987: Origins and levels of monthly and seasonal forecast skill for United
States surface air temperatures determined by canonical correlation analysis. Mon. Wea. Rev., 115, 1825-1850.
Barnston, A.G., 1994: Linear statistical short-term climate predictive skill in the Northern Hemisphere. J. Climate,
5, 1514-1564.
Barnston, A.G., W. Thiao and V. Kumar, 1996: Long-lead forecasts of seasonal precipitation in Africa using CCA.
Wea. Forecasting, 11, 506-520.
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. The CCA-based rainfall probability anomaly forecasts for the southern African region for Jan-Feb-Mar
1997, made at 1 month lead. Values shown are deviations (X100) 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.
