Spatial teleconnections in African rainfall: A comparison of 19th and 20th century patterns

Abstract

Rainfall variability over the African continent is remarkably coherent on large-spatial scales. The continent as a whole has exhibited marked fluctuations of rainfall on scales from decades to millennia. A newly created data set commencing in 1800 allows for the comparison between patterns of spatial variability that prevailed on modern and historical time scales. The results show the dominance of a small number of spatial patterns, one of the most common being anomalies of the same sign over most of the continent. Two other common modes are an opposition between equatorial and subtropical regions and an east–west opposition in the equatorial region and in southern Africa. In these modes, the shift between negative and positive poles can be abrupt. These patterns remain stable over time. That is, the spatial patterns and their degree of dominance are markedly similar during the 19th and 20th centuries. On the other hand, the large-scale factors governing the rainfall regime appear to vary on decadal time scales. This is particularly true for ENSO and the Indian Ocean Zonal Mode. The associations with large-scale wind regimes are more stable over time. The robustness of these patterns and varying importance of large-scale factors have implications for the interpretation of paleoclimate data.

Keywords

African rainfall climate change rainfall variability teleconnections

Introduction

One of the remarkable characteristics of African rainfall variability is its continental-scale coherence (Nicholson, 1986). On both annual and decadal scales, the most frequent mode of variability is anomalies of the same sign over most of the continent. A second major mode is one in which anomalies in the equatorial region are out-of-phase with those over most of the rest of the continent, particularly the subtropical regions of both hemispheres. Continental-scale variability is also evident in projections of climate change in response to greenhouse warming (Cook and Vizy, 2012).

A recently created precipitation data set for the 19th century permits an examination of the stability of these modes of variability and their associations with large-scale atmospheric parameters (Nicholson et al., 2012a, 2012b). The principal components are calculated for the continent as a whole and at a regional scale for northern- and southern-hemisphere sectors and for the equatorial sector of Africa. The data sets and methodology are described in the section ‘Data and methodology’. The continental-scale and regional-scale principal components are presented in the sections ‘Results for the 20th century’ and ‘Results for the 19th century’, respectively. Section ‘Teleconnections to the large scale: sea-surface temperatures and tropical modes of variability’ examines the frequency at which these modes appear in the modern and the historical records. Section ‘Teleconnections to the large scale: atmospheric factors’ presents select examples of changes in rainfall characteristics or teleconnections to the large scale that can be identified because of the availability of these long, historical records. Section ‘The seasonal cycle’ summarizes the results and presents the implications of the results for interpreting paleoclimatic data.

Data and methodology

Modern gauge data

The author has compiled a monthly gauge data set for the African continent that includes well over a thousand stations (Nicholson, 1986). The selection of stations for this archive was based on length of record and spatial distribution. Most records commence prior to 1925 and extend to 1998.

The individual gauge records are spatially averaged to produce time series for the 90 regions shown in Figure 1. Annual precipitation is expressed as a standardized departure from the mean, averaged over all stations in the region in the year in question. This approach is described in Nicholson (1986) and is commonly used in the analysis of African precipitation (e.g. Ali and Lebel, 2009).

Figure 1.

African regions used in the analyses and example of wetness classes averaged for the period 1825–1834 anomalies.

The 19th-century data set

A semi-quantitative precipitation data set covering the 19th century has been produced for the regions shown in Figure 1. The development of this data set is described in detail in Nicholson et al. (2012a, 2012b) and Lebel and LeBarbé (1997). It combines documentary information, hydrological indicators and all available gauge data. Data are expressed in terms of a seven-class ‘wetness’ index indicating conditions ranging from extremely wet to extremely dry, with a zero value indicating near-normal conditions. The temporal resolution is 1 year. Figure 1 shows an example, with classes averaged for the time interval 1925–1834.

To extend the coverage to the continent as a whole, two statistical steps were taken to fill in gaps in the record (Nicholson, 2000; Nicholson et al., 2012a, 2012b). The first involves statistical inference, based on linear correlation among regions. Most regions have a set of possible ‘substitute’ regions, with which the correlation is very high (exceeds the 0.001% level). The second was to fill in any remaining gaps utilizing a standard method of climate reconstruction (e.g. Mann and Jones, 2003; Neukom et al., 2010) based on principal component analysis. A validation of the data set suggested that the standard error is less than one anomaly class and the error in anomaly sign is less than 10% (Nicholson et al., 2012b).

Figure 2 shows the type of data used in the reconstruction for each year of the 19th century. Prior to 1820, roughly three-quarters of the regional values were obtained via climatic reconstruction. Hence, the reliability is in question during those first two decades. After 1890, the majority of regional values are derived from gauge data.

Figure 2.

Matrix of sources for the African climate reconstruction. Shadings distinguish between gauge data, other direct information (documentary and hydrologic), regional substitutions, and spatial reconstruction (from Nicholson et al., 2012b).

Methodology

Principal component analysis has long been used in meteorology to catalog map types or patterns (e.g. Lorenz, 1956). It reduces the variance in a data set by identifying the most common modes of variability. Here, it is used to identify dominant spatial patterns of rainfall variability and to quantify their temporal occurrence.

Input to the analysis in an N × M matrix of rainfall anomalies, where N is the number of regions and M is the number of years. In this study, a covariance matrix is used because it yields principal components that emphasize regions of greatest temporal variability. The analysis is S-mode, so that the eigenvectors are spatial patterns of rainfall anomalies and the principal components are time series of their coefficients (Richman, 1981). Compagnucci and Richman (2008) have shown that S-mode results well represent spatial teleconnections. Unrotated PCs are considered, in order to emphasize spatial teleconnections, as opposed to regional concentrations of variability.

The analysis was carried out for the continent as a whole and for the three overlapping sectors. The northern sector includes 41 regions, the equatorial region includes 42 regions, and the southern sector includes 51 regions. In order to compare and simultaneously analyze the historical data set and the modern gauge data, the latter were converted to the seven-class system, following Nicholson et al. (2012a). The correspondence between the classes and standardized anomalies is shown in Figure 3 for the Sahel region of West Africa, which roughly lies between 12° and 18° of latitude. The two series are extremely similar. Each PC analysis for modern data was carried out using both class values and standardized anomalies. The results were nearly identical in both loadings and spatial patterns.

Figure 3.

Correspondence between Sahel annual rainfall anomalies expressed as normalized departures and as converted to the seven-class system.

Several approaches have been utilized to determine how many principal components are significant and likely physically realistic (e.g. Kaiser, 1958; Kendall, 1980; North et al., 1982; Overland and Preisendorfer, 1982). Kendall’s (1980) criterion is based upon the need for the sampling error of an eigenvalue associated with a sample size N being smaller than its spacing from neighboring eigenvalues. Kaiser (1958) suggests that any PC with an eigenvalue greater than unity explains more variance than white noise. These two criteria suggest that as many as 10 PCs may be significant. However, the coherence of the spatial patterns, the spatial loadings, and comparison with known modes of variability suggest that in the current analyses, only the first three or four PCs are meaningful representations of large-scale patterns of variability.

A further issue that had to be resolved was related to the diversity of the seasonal cycle over the continent. In most of northern-hemisphere Africa, the driest season is the boreal winter. In this case, a calendar-year calculation of annual totals is appropriate. The calendar year is also appropriate in the equatorial regions, which have two rainy seasons that occur during the transition seasons. In most of Southern-Hemisphere Africa, where the rainy season falls in the austral summer, the calendar year bisects the rainy season, and ‘rain year’ or ‘agricultural year’ totals, generally starting in July, are frequently utilized.

This suggests it may be problematic to evaluate annual data for the continent as a whole. For this reason, two preliminary analyses were carried out – one using calendar-year totals everywhere and the second using July-to-June totals for areas with peak rainfall in the austral summer and calendar-year totals elsewhere. The results of both analyses were similar, largely because most of the interannual variability in southern Africa is associated with March-to-May rainfall. Thus, the principal component analyses presented here all utilize calendar-year rainfall.

The modern PC analysis is confined to the years 1920–1994, in order to reduce the number of missing data points. During this time period, less than 3% of the regional values were missing, and these were assumed to be zero (i.e. average rainfall). The historical analysis was carried out for the years 1820–1900. The earliest two decades of the century are omitted because they are based largely on climatic reconstruction (Nicholson et al., 2012b). Because the sign convention for the PCs is arbitrary, the chosen representation of each PC is that which highlights negative loadings.

An initial goal was to determine whether the large-scale links to the dominant PCs (e.g. SST patterns) differed for the historical and modern periods. However, an examination of PC time series indicated that the region with the strongest correlation to the PC dominated the associations with the large scale. For this reason, regional rainfall series were used instead to produce select examples of historical versus modern links to the large scale. The examples chosen include the July-to-September season for the Sahel and the October–November season for East Africa.

SST data were obtained from the Extended Reconstruction SST data set (ERSST), which has 2-degree resolution and extends back to 1854 (Kaplan et al., 1998; Smith et al., 2008). Zonal winds at the surface and 200 mb were respectively taken from the ICOADS marine data set (Woodruff et al., 2011) and the 20th-Century Reanalysis V2 data set (Compo et al., 2011). Niño 3.4 was used as an index of the ENSO phenomenon, and the Indian Ocean Zonal Mode (IOZM) index, a measure of east–west temperature contrast in the equatorial Indian Ocean, was based on the ERSST data set.

Results for the 20th century

Continental-scale modes of variability

Figure 4a (left) shows the eigenvectors corresponding to the first four principal components of modern continental-scale variability. The first PC accounts for 16% of the interannual variability over the 75-year period, and the four collectively account for 40% of the variability. The first shows a pattern in which anomalies are in-phase throughout most of the continent, but they are much weaker in the equatorial region. The eigenvector for the second PC shows an east–west opposition in the low latitudes as well as an out-of-phase relationship between subtropical regions of the two hemispheres. The eigenvector corresponding to the third PC, like the first, shows distinct contrasts between the equatorial region and the subtropics. However, the loadings are very weak in most of the northern hemisphere. The spatial patterns associated with PCs 1 and 3, in-phase anomalies throughout the continent and strong opposition between equatorial and subtropical regions, have long been recognized as dominant modes of African rainfall variability on both annual and decadal time scales (Nicholson, 1986, 2000). The eigenvector associated with the fourth PC shows an interesting pattern of positive anomalies in the Atlantic coastal sectors (except in the western Sahel) and in parts of northern-most Africa, and strong negative anomalies most elsewhere. This is suggestive of a strong localized link to the Atlantic Ocean.

Figure 4.

(a) Continental-scale eigenvectors: left – for the period 1920–1994; right – for the period 1820–1900. Loadings are indicated for each region, and the percentage variance explained by each EOF is indicated. (b) Eigenvectors based regions of northern Africa: left – for the period 1920–1994; right – for the period 1820–1900. Loadings are indicated for each region and the percentage variance explained by each EOF is indicated. (c) Eigenvectors based regions of equatorial Africa: left – for the period 1920–1994; right – for the period 1820–1900. Loadings are indicated for each region and the percentage variance explained by each EOF is indicated. (d) Eigenvectors based regions of southern Africa: left – for the period 1920–1994; right – for the period 1820–1900. Loadings are indicated for each region and the percentage variance explained by each EOF is indicated.

Regional-scale variability

Both of the aforementioned patterns are even more strongly evident in the analysis for northern Africa (Figure 4b, left). The eigenvector of the first PC shows positive anomalies in nearly every region. The next three all show an opposition between equatorial and subtropical latitudes. These differ mainly with respect to the east–west extent of the equatorial region that is out-of-phase with the subtropical latitudes. In the second PC, the opposition is confined to the central and western sectors. An interesting feature is anomalies of the same sign in the Guinea Coast and from the central Sahara northward. For northern Africa, the first PC accounts for 23% of the variance and the first four collectively account for 48%.

The results for the equatorial region itself show three distinct patterns (Figure 4c, left). The eigenvector for the first PC shows in-phase anomalies over most of the area from 10°N to 15°S, but with very strong loading in the eastern sector. That PC accounts for 19% of the variance. The second and third, respectively, show an east–west opposition and the opposition between equatorial and higher latitudes mentioned earlier. In this latter case, the contrast is at roughly 5°N and 10°S. These two PCs account for another 21% of the variance, giving a total of 40% of the variance explained by three PCs.

The PCs for the southern hemisphere (Figure 4d, left) further emphasize the opposition between equatorial and subtropical regions. This is evident in the first two, which differ mainly in the spatial extent of the lower latitude anomaly. The first, which explains 28% of the variance, has strong and in-phase anomalies from c. 10°S to the southern extreme of the continent. The loadings are weak in the equatorial region. The second PC, which accounts for 11% of the variance, has the strongest loadings in the east, and anomalies are strong in both the equatorial and subtropical regions. The node in the pattern is at roughly 15°S. The dominant feature of the third PC is an east–west opposition, although anomalies in the eastern parts of the subcontinent tend to be of opposite sign north and south of c. 25°S. PCs 1 to 3 collectively explain 48% of the variance.

Results for the 19th century

Figure 4a (right) shows the eigenvectors of the first four continental-scale PCs for the 19th century. These together account for 51% of the variance, with the first accounting for 20%. The first three show strong similarities to the eigenvectors of the modern PCs, especially PC1 and PC2. However, there are notable differences in the dominant loadings. In PC1, for example, strong anomalies of a single sign are predominant over most of the continent. In the modern case, there is a narrow equatorial area with anomalies that are either extremely weak or opposite in sign to the prevailing continental pattern. Also, the dominant loadings are in southern Africa. In the historical case, the equatorial opposition is only a narrow strip along the Atlantic coast, and the dominant loadings are mainly in the Sahel. The eigenvector corresponding to the fourth historical PC shows the common modern anomaly pattern of a strong opposition between equatorial and subtropical latitudes. Although this pattern was not apparent in the first four PCs for the 20th century, it was the second most common mode during the period 1901–1973 (Nicholson, 1986).

Figures 4b–d (right) show the eigenvectors for the first three 19th-century PCs for North Africa, the equatorial region, and southern Africa, respectively. As in the continental-scale analysis, the similarity in pattern with the 20th-century PCs is striking in each case. The major modes illustrated by the first PCs reinforce the conclusion that the most important mode of variability is coherent fluctuations throughout most of the continent. However, PC1 for the equatorial region and PC1 for southern Africa show an opposition between coastal regions in proximity to the eastern equatorial Atlantic and remaining parts of the sector. The secondary modes are distinct for the three sectors. For North Africa, PCs 2 and 3 exhibit the well-known opposition between equatorial and subtropical latitudes. For equatorial Africa, PC3 shows this pattern to some extent, but an east–west opposition is dominant in PC2. For southern Africa, PC3 likewise shows an opposition between equatorial and subtropical latitudes. PC2 shows a strong contrast between a southern–eastern sector with strong positive anomalies and strong negative anomalies nearly everywhere else.

The similarity between the modern and historical eigenvectors is strong evidence that these spatial modes are inherent attributes of rainfall variability over Africa. They probably reflect fundamental forcing mechanisms at the continental and regional scale. Notably, the similarity cannot be attributed to the spatial teleconnections utilized in the climatic reconstruction for the 19th century, because the spatial scale of the teleconnections is much smaller than these analysis sectors.

Time series of the principal components

While the eigenvectors illustrate the spatial modes of variability, the corresponding principal component time series represent the degree to which a given spatial mode is developed in each year. For the sake of homogeneity, the principal components calculated from the modern data are utilized for both the modern and historical periods. Thus, principal component loadings for each mode will be calculated for every year from 1820 to 1994. This approach is justified by the similarity of the modern and historical PCs and will permit a direct comparison of the frequency of the modes across the two centuries. In each case, only the first two PCs will be evaluated.

Figure 5 shows the PC time series for PCs 1 and 2 for the continental-scale analysis. PC1, with anomalies of the same sign throughout most of Africa, was very prominent throughout the 19th century. A period of dry conditions prevailed throughout the 1820s and 1830s, with a return to dry conditions in the late 1840s. From the early 1850s onward until 1902, the wet phase of this mode prevailed almost uninterrupted. This pattern of pronounced low-frequency variability was replaced by interannual variability in the 1920s through 1940s. A wet phase prevailed throughout the 1950s. The intermittent positive development of this mode in the 1960s and 1970s was actually related to regional-scale anomalies in the equatorial region and the southern-hemisphere region, respectively. The long dry phase commencing in 1979 was apparent primarily in the North African and southern-hemisphere regions.

Figure 5.

Principal component time series for first and second EOFs for each sector. The dashed black line indicates the trend from 1840 to 1994.

The second principal component, characterized by an opposition between the northern and southern-hemisphere sectors, alternated between its negative and positive phases throughout most of the 19th century. During the 20th century, the negative phase prevailed from the 1920s until the 1970s, roughly the onset of the severe and multi-decadal drought in the Sahel and surrounding regions of North Africa.

Both PC1 and PC2 are indicative of increasing aridity since the mid-19th century. However, PC2 reflects mainly variations in the Sahel; its correlation with PC1 of the North African analysis is −0.89. PC1 for the continent and PC1 for North Africa are also well-correlated: r = 0.50. The main contrast between them is the degree of interannual persistence; low-frequency variability is much more pronounced in the latter. PC1 is characterized by anomalies of the same sign throughout most of the analysis sector, but the anomalies are clearly largest in the Sahel. Its sign was positive in 80 of the 104 years from 1866 to 1969. It was negative in all but two of the 24 years from 1970 to 1993. PC2, characterized by an opposition between the Sahel and Guinea Coast, shows little evidence of such low-frequency variability.

PC1 for equatorial Africa shows a high degree of interannual persistence in the 19th century, but strong interannual variability in the 20th century. This tendency may relate in part to the nature of the proxy data used in the 19th century. However, similar contrast between the two centuries is not apparent in PC2, suggesting that the 19th-century low-frequency variability is real.

For southern Africa, PC1 similarly shows a tendency for increased variability in the 20th century, compared with the 19th. This same tendency is evident in PC2. The low-frequency variability appears to be in-phase in the two PCs. However, the interannual variability appears to be largely out-of-phase during the 20th century, when high-frequency variability prevailed.

Each of the PCs shown in Figure 5 was subjected to a trend analysis, based on least squares regression. In one case, the trend calculation commenced at the beginning of the PC series, that is, in 1820. Because of uncertainty in the earliest years of the analysis, a second trend calculation commenced in 1840. Only the latter set of trend lines is shown, because the longer term trends are near zero and thus generally indistinguishable from the x-axis. Clearly, no significant trend is apparent in any of these cases. This implies a very long-term stability of the precipitation regime over Africa.

Teleconnections to the large scale: Sea-surface temperatures and tropical modes of variability

The principal components of rainfall variability in the four analysis regions are markedly similar in magnitude and spatial pattern during the historical and modern periods. Some of the similarity may be a result of the methodology of the spatial reconstruction, that is, filling in gaps based on correlations with nearby regions. However, the spatial scale of the correlations utilized in the reconstruction is much smaller than the spatial coherence exhibited by the PCs. Also, this methodology cannot account at all for the similar dominance of the various spatial modes in the historical and modern periods. The similarity of the spatial patterns of variability in the 19th and 20th centuries and the degree of dominance of the various patterns suggest that common factors in variability prevailed in both the historical and modern periods.

This assertion is tested by examining the relationship between rainfall variability and select large-scale phenomena. Two regions, the Sahel region and East Africa, are chosen for these analyses because they provide most of the variance in the first PCs for the northern and equatorial analyses. Their locations are shown in Figure 6. Time series of rainfall variability in these regions is correlated with sea-surface temperatures. In section ‘The seasonal cycle’, correlations are carried out with atmospheric circulation. The rainfall time series, as opposed to the PCs, are used because they represent somewhat more coherent sectors than do the PCs and because this allows for distinguishing between the long and short rains of East Africa (March-to-May and October–November, respectively). The results for the long rains are not considered here because teleconnections to the large scale were shown to be very weak, in agreement with the conclusions of other studies of this season (e.g. Camberlin and Okoola, 2003; Hastenrath et al., 2011; Mutai and Ward, 2000).

Figure 6.

Location of sea-surface temperature and wind sectors and geographical regions used in the study.

To determine the stability of the teleconnections between rainfall and potential large-scale drivers, linear correlation coefficients are calculated for running 20-year periods, starting with 1874, the first year of the East African series. The SST sectors to be correlated with rainfall were determined by first producing maps of the correlation with global SSTs between 75°N and 75°S. SST sectors best correlated with rainfall were subjectively delineated.

The so-identified sectors are shown in Figure 6. For the Sahel, these include the north equatorial Atlantic, the equatorial Indian Ocean, the Mediterranean Sea, and a sector in the northern Pacific. The Sahel rainfall index is a standardized departure for the July-to-September season. The correlation is with simultaneous variables. For the short rains of East Africa, that is, the October–November season, the equatorial Indian Ocean and two areas of the south-equatorial section of the western Pacific were selected, along with the IOZM. For the long rains, correlations with global SSTs were generally low, so that no sectors were identified for further analysis. The rainfall series were also correlated with sea-surface temperature in the Niño 3.4 region of the equatorial Pacific during both the concurrent and previous 3-month seasons.

The running 20-year correlations for the Sahel are presented in Figure 7a. The magnitude and even the sign of correlations change over time, even within the modern period. Those for the Atlantic and Indian Ocean are generally insignificant before the 20-year segment beginning about 1920 but are significant and negative from then to around 1970. The temporal sequence of correlation with the Mediterranean is very different, being positive and significant from early in the 20th century to the 20-year period beginning around 1950. The temporal sequence of correlation with the northern Pacific sector is strikingly similar to that of the Mediterranean.

Figure 7.

(a) Twenty-year running correlations between rainfall and SSTs in various sectors: top – annual Sahel rainfall versus SST; bottom – East African October–November rainfall versus SSTs (see Figure 6 for location of SST sectors). (b) 20-year running correlations between rainfall and zonal winds for the Sahel rainy season (July-to-September) and for East African October–November rainfall.

The correlation of the East African short rains with large-scale indices is shown in Figure 7a. Rainfall tends to be positively correlated with Niño 3.4, the IOZM, and SSTs in the equatorial Indian Ocean, but negatively correlated with subtropical SSTs in the central Pacific. However, the correlations with three of the four indicators become transiently insignificant commencing with the 20-year interval starting in 1899. That with the IOZM remaining weak over roughly four decades, ENSO appears to be a stronger driver than the Indian Ocean during that interval. Notably, October–November rainfall is abnormally low during that time (Nicholson, forthcoming).

Teleconnections to the large scale: Atmospheric factors

For the Sahel, the atmospheric factor most closely linked to interannual variability is the strength of the Tropical Easterly Jet over West Africa (Nicholson and Grist, 2001). For the short rains of East Africa, it is the low-level zonal wind over the central equatorial Indian Ocean (Hastenrath et al., 2011), but the link to 200 mb winds in the same region is also exceedingly strong (Nicholson, forthcoming).

Figure 7b shows running 20-year correlations between Sahel rainfall in JAS and 200 mb zonal winds near the core of the Tropical Easterly Jet over West Africa. The overall correlation between values of the two parameters in individual years is −0.39 over the 128-year period. This is consistent with studies such as Nicholson and Grist (2001), Nicholson (2009), and Fontaine et al. (2011), showing the overwhelming importance of the Tropical Easterly Jet as a factor in Sahel rainfall.

Figure 7b also shows running 20-year correlations between October–November rainfall and zonal winds at 1000 and 200 mb between 5°N and 5°S in the central Indian Ocean. The long-term correlation between rainfall and the surface zonal wind over the period 1874–1996 is −0.61. The 20-year running correlations range from −0.25 to −0.90, but remain above the 10% significance level (0.38) during most of the time period commencing in the 1870s and ending in the 1980s. For 200 mb zonal winds, the relationship to rainfall is weaker but more consistent, remaining significant throughout most of the analysis period. Except for roughly one decade in the late 19th century, when the reliability of the 200 mb data is questionable, the correlation ranged from 0.35 to over 0.8.

These results demonstrate that the relationship to the zonal winds over the central equatorial Indian Ocean is not constant, but that it is evident in both the historical and modern periods. Notably, during the period when the relationship is weak, from the late 1890s to the 1940s, the correlation with ENSO (i.e. with Niño 3.4) is high, and the equatorial westerlies are relatively weak (Nicholson, 2014). The result of these two opposing effects is a period of relative constancy of the October–November rainfall.

The seasonal cycle

Modern analyses have shown several features of the seasonal cycle over the Sahel and East Africa and its relationship to interannual variability. For the Sahel, it has generally been noted that most of the interannual variability is linked to changes in August and September rainfall (e.g. Dennett et al., 1985; Nicholson and Palao, 1993). For East Africa, numerous studies have shown that the ‘short rains’ of October–November, the secondary rainy season, account for most of the interannual variability. The historical time series allows the temporal stability of these conclusions to be examined.

Figure 8 shows the sliding 20-year correlations between annual rainfall in the Sahel and rainfall in three seasons: April–May, August–September, and October–November. In the last decades of the 19th century, both April–May and October–November were strongly correlated with annual rainfall, while the correlation with August–September was very weak. Throughout most of the 20th century, the highest correlation has been with August–September, with relatively little contribution from April–May and October–November. The shift in the importance of these months is equivalent to a shift between annual totals being linked to changes in the onset or end of the season versus the amount of rainfall during the peak months of the season. Presumably different large-scale factors govern the occurrence of these two cases.

Figure 8.

Twenty-year running correlations between seasonal and annual rainfall for the Sahel and East Africa.

Figure 8 also shows the sliding 20-year correlations between annual rainfall in East Africa and the October–November and March-to-May seasons. From the intervals beginning in the 1930s to present, the October–November season is strongly correlated with annual totals, comprising roughly 40–70% of the variance. The March-to-May season is strongly correlated with annual totals throughout the analysis period, except for the earliest years. The low correlation at that time relates to the major October–November rainfall event in 1877, coupled with both an extreme El Niño and an IOZM event in the Indian Ocean.

Summary and conclusion

The spatial teleconnections in rainfall over the African continent are remarkably stable over time. The most prominent patterns of variability are common to both the 19th and 20th centuries. Some of the most common modes are anomalies of the same sign over most of the continent, an opposition between equatorial and subtropical regions, and an east–west opposition in the equatorial region and southern Africa. In these modes, the shift between negative and positive poles can be abrupt.

A notable characteristic for each of the four sectors evaluated is extended aridity early in the 19th century. Although the quantitative indicators are scarce, proxy indicators confirm the extreme, protracted, and widespread nature of this interval (e.g. Bessems et al., 2008; Nicholson et al., 2012b; Verschuren et al., 2000). The core of this interval was the period 1825–1834, a decade in which strong negative anomalies prevailed over most of the continent (Figure 1).

Despite the occurrence of this interval, there have been no significant trends in the principal components over the last two centuries. This suggests a climatic regime that is very stable over the long-term, despite considerable interdecadal variability.

The spatial teleconnections demonstrated here have important implications for the interpretation of paleoclimatic indicators, especially lakes. A comparison of lake location with the principal component loadings shows that the Rift Valley lakes tend to lie along the node in the pattern of east–west opposition in PC2 (Figure 4c) and Lake Malawi, spanning the latitudes 9°S to 145°S, lies along the node of the north-south opposition in PC2 in Figure 4d. Lake Bosumtwi at 6.5°N is often cited as an indicator of Sahel climate (e.g. Shanahan et al., 2009), but it lies well within the region of opposition with the Sahel, as shown in PC2 in Figure 4b.

In contrast to the spatial teleconnections, the factors governing interannual variability vary on decadal time scales. This is particularly true for the associations with ENSO and the IOZM. The associations with the wind regimes, the more direct factors in variability, are more stable over time. This suggests that emphasis should be placed on how large-scale factors modify the regional circulation systems.

Footnotes

Funding

The author would like to acknowledge the support of two historical grants from NOAA: NAO80AR4310731 and A46GP0285.

References

Ali

Lebel

(2009) The Sahelian standardized rainfall index revisited. International Journal of Climatology 29: 1705–1714.

Bessems

Verschuren

Russell

. (2008) Paleolimnological evidence for widespread late 18th century drought across equatorial East Africa. Palaeogeography, Palaeoclimatology, Palaeoecology 259: 107–120.

Camberlin

Okoola

(2003) The onset and cessation of the long rains in eastern Africa and their annual variability. Theoretical and Applied Climatology 75: 43–47.

Compagnucci

Richman

(2008) Can principal component analysis provide atmospheric circulation or teleconnection patterns? International Journal of Climatology 28: 703–726.

Compo

Whitaker

Sardeshmukh

. (2011) The Twentieth Century Reanalysis Project. Quarterly Journal of the Royal Meteorological Society 137: 1–28.

Cook

Vizy

(2012) Impact of climate change on mid-twenty first century growing seasons in Africa. Climate Dynamics 39: 2937–2955.

Dennett

Elston

Rodgers

(1985) A reappraisal of rainfall trends in the Sahel. Journal of Climatology 5: 353–361.

Fontaine

Roucou

Gaetani

. (2011) Recent changes in precipitation, ITCZ convection and northern tropical circulation over North Africa (1979-2007). International Journal of Climatology 31: 633–648.

Hastenrath

Polzin

Mutai

(2011) Circulation mechanisms of Kenya rainfall anomalies. Journal of Climate 24: 404–412.

10.

Kaiser

(1958) The Varimax criterion for analytic rotation in factor analysis. Psychometrika 23: 187–200.

11.

Kaplan

Cane

Kushnir

. (1998) Analyses of global sea surface temperature 1856-1991. Journal of Geophysical Research-Oceans 103: 18567–18589.

12.

Kendall

(1980) Multivariate Analysis. London: Charles Griffin.

13.

Lebel

LeBarbé

(1997) Rainfall monitoring during HAPEX-Sahel. 2. Point and areal estimation at the event and seasonal scales. Journal of Hydrology 189: 97–122.

14.

Lorenz

(1956) Empirical Orthogonal Functions and Statistical Weather Prediction. Technical report, Statistical Forecast Project Report 1, Department of Meteorology, MIT (NTIS AD 110268).

15.

Mann

Jones

(2003) Global surface temperatures over the past two millennia. Geophysical Research Letters 30: 1820–1823. DOI: 10.1029/2003GL017814.

16.

Mutai

Ward

(2000) East African rainfall and the tropical circulation/convection on intraseasonal to interannual timescales. Journal of Climate 13: 3915–3939.

17.

Neukom

Luterbacher

Villalba

. (2010) Multiproxy summer and winter surface air temperature field reconstructions for southern South America covering the past centuries. Climate Dynamics 37: 35–51. DOI: 10.1007/s00382-010-0793-3.

18.

Nicholson

(1986) The spatial coherence of African rainfall anomalies. Interhemispheric teleconnections. Journal of Climate and Applied Meteorology 25: 1365–1381.

19.

Nicholson

(2000) The nature of rainfall variability over Africa on time scales of decades to millennia. Global and Planetary Change Letters 26: 37–158.

20.

Nicholson

(2009) On the factors modulating the intensity of the tropical rainbelt over West Africa. International Journal of Climatology 29: 673–689.

21.

Nicholson

Grist

(2001) A simple conceptual model for understanding rainfall variability in the West African Sahel on interannual and interdecadal time scales. International Journal of Climatology 21: 1733–1757.

22.

Nicholson

Palao

(1993) A reevaluation of rainfall variability in the Sahel. 1. Characteristics of rainfall fluctuations. International Journal of Climatology 13: 371–389.

23.

Nicholson

Dezfuli

Klotter

(2012a) A two-century precipitation dataset for the continent of Africa. Bulletin of the American Meteorological Society 93: 1219–1231.

24.

Nicholson

Klotter

Dezfuli

(2012b) Spatial reconstruction of semi-quantitative precipitation fields over Africa during the nineteenth century from documentary evidence and gauge data. Quaternary Research 78: 13–23.

25.

North

Bell

Calahan

(1982) Sampling errors in the estimation of empirical orthogonal functions. Monthly Weather Review 110: 699–706.

26.

Overland

Preisendorfer

(1982) A significance test for principal components applied to a cyclone climatology. Monthly Weather Review 110: 1–4.

27.

Richman

(1981) Obliquely rotated principal components: An improved meteorological map typing technique? Journal of Applied Meteorology 20: 1145–1159.

28.

Shanahan

Overpeck

Anchukaitis

. (2009) Atlantic forcing of persistent drought in West Africa. Science 324: 377–380.

29.

Smith

Reynolds

Peterson

. (2008) Improvements to NOAA’s historical merged land-ocean surface temperature analysis (1880 to 2006). Journal of Climate 21: 2283–2296.

30.

Verschuren

Lair

Cumming

(2000) Rainfall and drought in equatorial east Africa during the past 1,100 years. Nature 403: 410–414.

31.

Woodruff

Worley

Lubker

. (2011) ICOADS Release 2.5: Extensions and enhancements to the surface marine meteorological archive. International Journal of Climatology 31: 951–967.