extremes and at higher elevations. These are minor compared to rainfall fluctuations; moreover, they have not been systematically studied. Therefore, this paper will be limited to variability of rainfall.
Most of the description of rainfall variability in this paper derives from previous studies made by the author. Analyses are based on a network (Figure 1) of approximately 1400 stations, the statistics of which are fully described in Nicholson et al. (1988). From these, a spatially averaged data set has been derived, utilizing 90 geographical regions (Figure 1) that are homogeneous with respect to the interannual variability of rainfall.
Regional averages are generally expressed as a standardized annual departure from the long-term mean (i.e., the mean over the entire length of record). In select analyses, the rainfall anomalies are also presented as a percentage departure from the long-term mean. About 75 percent of the station records commence before 1925 and continue to 1984 or later. Hence, the length of record is generally at least 60 years.
Regionally averaged rainfall Rj is calculated as
where Rj is the regional rainfall departure in the year j, and Ij is the number of stations available in year j. In the above, xij is a standardized rainfall departure at station i in
year j. It is calculated as
where rij is the annual total at the station in the year j, is the station's long-term annual mean, and si is the variance of annual rainfall at the station.
These regional means are utilized to examine the temporal and spatial structure of rainfall variability over the continent. They also illustrate the magnitude of decadal anomalies. In some analyses 1°-grid averages are utilized instead of the regional averages.
Previous studies (e.g., Nicholson, 1986) have utilized a map-classification technique of Lund (1963) to describe the preferred spatial configurations of rainfall anomalies over Africa. This technique is based on linear correlation of map patterns. Most of the rainfall variability over the continent is described by the six "anomaly types" illustrated in Figure 2. Unlike eigenvector techniques, the patterns produced by the Lund method are not orthogonal; hence, these six anomaly types really describe only four basic configurations of anomalies. These include a tendency for above(or below-) average rainfall over most of the continent (Types 2 and 4) and two patterns (Types 1, 3, 5, and 6) illustrating an opposition between equatorial and subtropical latitudes. Approximately 21 of the 90 years between 1901 and 1990 are represented by a pattern of above-average rainfall in the equatorial regions but below-average rainfall in subtropical latitudes, and 12 by the opposite pattern, which is illustrated by Type 3 in Figure 2. Fourteen of the 70 years are best represented by the Type 2 anomaly pattern, with below-average rainfall throughout most of the continent, and 15 by Type 4, with above-average rainfall in most areas.
These patterns are introduced here for two reasons. First, as will later be seen, they are typical of decadal-scale rainfall fluctuations over Africa. Second, they demonstrate that much of the rainfall variability over the continent can be described using a small number of regional time series. The representative regions, shown in Figure 3, include five latitudinal zones running from the southern margin of the Sahara to the Guinea Coast of the Atlantic, eastern equatorial Africa, and two sectors of southern Africa loosely termed the northern and southern Kalahari.
Two major, decadal-scale rainfall anomalies have affected the African continent within the current century: