The economy of predominantly rural southern Africa is largely agro-based and the region's agriculture is principally rain fed. The precipitation shows high variability on all timescales (Mason and Jury, 1997; Tyson and Preston-Whyte, 2000; Reason and Rouault, 2002) making the region well known for devastating rainfall extremes. The most recent examples of such extremes are the severe droughts of 1991/1992, 2002/2003, and 2003/2004 that affected most of southern Africa and the major flooding episodes that devastated Mozambique during 2000 and 2001 in which hundreds of people were killed while another 200 000 were made homeless (Reason and Keibel, 2004; Washington and Preston, 2006). Consequently, strong challenges exist for sustainable development in this part of the world. Thus, clarifying the rainfall variability of the region and its associated influencing factors are of great importance.
Although the underlying mechanisms behind the regional climate variability are not well understood, better understanding is not only important for climate predictions, but also for historical climate attribution and realistic future climate change projections. A review of rainfall mechanisms and triggers over southern Africa was presented by Mason and Jury (1997). In this, and other earlier studies, focus was placed on ENSO as the major cause of Southern African rainfall variability (e.g. Ropelewski and Halpert, 1987; Janowiak, 1988; Lindesay, 1988). Lindesay (1988) for example, found that stronger correlations with ENSO exist later in the rainfall season (January–March) than in the earlier part (October–December). In support, several model results also suggest that drying over southern Africa can occur mainly as a consequence of the atmosphere's response to ENSO warming in the eastern Pacific (Rocha and Simmonds, 1997).
Even when equipped with this accumulated knowledge of ENSO regional impacts and using this climate mode as the principal predictor, forecasting seasonal rainfall over the region remains complex and poorly resolved. In fact, other observational studies suggest that the ENSO signal is not very strong in southern Africa, and that regional interannual rainfall variability is largely forced by the SST anomalies in the Indian and/or southern Atlantic Ocean that may not be causally connected to the ENSO cycle (e.g. Mason, 1995; Nicholson and Kim, 1997). These observational studies were supported by related model outputs which suggest little evidence or even the absence of ENSO signal over the region (Goddard and Graham, 1999). Other authors have proposed that the ENSO signal propagates into the southern African precipitation field from more than one source, increasing the complexity in seasonal rainfall prediction (e.g. Cook, 1998; Reason and Jagadheesha, 2005).
More recently, the identification of the IODZM (Saji et al., 1999; Webster et al., 1999) as a climate mode in the tropical Indian Ocean that seems to be mainly independent of ENSO, led to studies that provided evidence supporting the significant role of this mode in influencing African rainfall variability during the 20th century (e.g. Black et al., 2003; Saji and Yamagata et al., 2003). Manatsa et al., (2007, 2008) showed that the IODZM could have provided a potentially more robust rainfall predictor for Zimbabwe summer rainfall than ENSO for the last half of the 20th century. In addition to the IODZM, a dipole SST pattern in the south Indian Ocean may also be important to the late summer rainfall of the eastern areas of southern Africa (Behera and Yamagata, 2001; Reason, 2001). This mode which was coined as the South Indian Ocean dipole (SIOD) manifests itself as warm (cool) and cool (warm) anomalies in the western and eastern subtropical South Indian Ocean respectively. Despite ongoing research efforts, it appears that southern African rainfall variability is still not sufficiently well understood to enable significant improvements in seasonal forecasting for the region.
An aspect of many previous climate studies is that most researchers seem to have assumed stationarity in the relationships during their analyses and in building prediction models. In fact, potential rainfall predictors may experience significant interdecadal variations, especially since southern Africa is known to exhibit strong decadal signals in the observational record (Reason and Rouault, 2002). In addition, there appear to be decadal variability in the ENSO-IODZM relationship (e.g. Schott et al., 2009) and the impact of the IODZM on southern African rainfall (Manatsa and Matarira, 2009).
Since these potential predictors could be quite significant during some decades and not in the others (e.g. Kumar et al., 1999), the main task of this study is to examine in more detail the changing IODZM rainfall relationship in the context of the decadal shifts in the Indian Ocean climate system. The emphasis is on establishing the period around 1997 as a potentially important turning point in the Indian Ocean climate system. Evidence on epochal shifts in the IODZM-southern African rainfall link is determined by focusing on the different relationship responses to the pre-1961/1962, 1961/1962 to 1976/1977, 1977/1978 to 1997/1998 and finally the post-1997/1998 period. The hypothesis is that although there may be other significant regime shifts, 1997/1998 (Manatsa and Matarira, 2009) is the major turning point in the Indian Ocean climate system whose extended impacts include the decoupling of the IODZM from the determinant mechanisms of the southern African early summer rainfall, with SST anomaly shifts in the subtropical south Indian Ocean also playing an important role.
In this study, focus is deliberately placed on the Indian Ocean variability, although we recognize that both the Pacific and Atlantic Ocean SST play important roles in southern African rainfall (e.g. Reason et al., 2006). Although the rainy season over most of southern Africa stretches from October to March, the current investigation only considers the rainfall anomalies for the early summer, October to December (OND). This differentiation of the rainy season as distinct early and later summer periods is also recognized by the southern African Regional Climate Outlook Forum (SARCOF) which issues seasonal rainfall forecasts separately for the two sub seasons. The rationale for studying only the OND period is that the triggering mechanisms which are mainly responsible for the early and later summer seasons are found to be fundamentally different (D'Abreton and Tyson, 1995). Furthermore, the early part coincides with the highly active period of the IODZM. Moreover, this OND period is less well understood than the JFM season, and its rainfall is less reliable and more variable than that for JFM (Clark et al., 2003; Richard et al., 2001). Thus, there are potentially great social and economic benefits if the understanding and forecasting of the OND season rainfall can be improved.
2. Data and methods
In this study, variability is used as the major descriptive parameter for observational time series (e.g. Manatsa and Matarira, 2009) to capture recent prominent climate shift periods. NCEP/NCAR reanalysis data, available for 1948 to the present at 2.5° horizontal resolution (Kalnay, 1996), were used to investigate climate variability in the region. Longer-term datasets from the IODZM index (Saji et al., 1999) and ENSO (Niño 3.4) index (Kaplan et al., 1998) and regional rainfall available from 1922 to 2008 were also used. All the analyses were performed on unfiltered data using linear statistical methods. In order to assess the interdecadal variations in the Indian Ocean SST and rainfall, overlapping periods of analysis were used. These periods are 31 years long so as to provide for the standard climatological normals of the various climate variables. Indices derived from the available variables (SSTs, MSLP and zonal and meridonial (u, v)-wind stress) are analysed to show that the turning points observed are not peculiar to one specific data set but are common to all.
The data were then examined in epochs whose decomposition was differentiated by the analysis of the nature of low frequency signal in the climatic record. We used the F-test to compare interannual variances between epochs and the t-test to assess the statistical significance of differences in the epoch means, with α set at 95 and 99% confidence levels. For wind stress, the zonal (u) and meridional (v) components were tested independently. The MSLP and SST anomaly maps for the epochs are drawn online using facilities from the Climate Prediction Centre website, http://www.cdc.noaa.gov/cgi-bin/Composites and Climate Explorer website, http://climexp.knmi.nl/. This approach was used to analyse the sea surface temperature, pressure and circulation variations associated with the 1997 interdecadal shift of the Indian Ocean climate system and to see how it relates to the decoupling of the IODZM from the southern African rainfall variability. For the IODZM, the IODZM index (Saji et al., 1999) which quantifies zonal gradients in the SST rather than a proper east-west dipole has been used. Thus, we opted to use the term IODZM (as used by Annamalai et al., 2004 and Clark et al., 2003) rather than IOD (e.g. used by Saji et al., 1999), as it matches the observed out of phase rather than the simultaneous variations in the development of the SST extremes in the east and west of the Indian Ocean.
Following Saji et al., (1999), we considered the IODZM index as the anomaly difference between the SST anomalies of the western (50°E–70°E, 10°S–10°N) and eastern (90°E–110°E, 10°S–equator) tropical Indian Ocean. The SSTs for the IODZM index dataset (from 1901 to 2008) were adopted from Kaplan et al., (1998) and extracted from http://www.jamstec.go.jp/frcgc/research/d1/iod/kaplan_sst_dmi.txt. The IODZM index is calculated from the average of the most active months of October to December (OND) and normalized relative to its standard deviation. We adopted ± 0.7 as the value above/below which defines the positive/negative IODZM events as shown with coloured bars in Figure 1. This value has been found as the threshold beyond which southern African rainfall (Manatsa et al., 2009) begins to show notable sensitivity to the IODZM index values. In any case, the resulting IODZM event years depicted in Figure 1 closely resemble those proposed by Behera et al., (2005).
To assess the strength of ENSO during the OND season, the familiar Niño 3.4 SST index (Ropelewski and Halpert, 1987) is used. This index is calculated by taking the average of SST anomalies over 5°N–5°S, 170°W–120°W and standardized using its standard deviation. The observed SST anomalies are computed from the Hadley Centre sea ice and SST (HadISST) data set (Rayner et al., 1996). OND is the season when the influence of ENSO has been found to be strong over the region as compared to other lagged periods (Manatsa et al., 2008). In Figure 1, anomaly values in the Niño 3.4 index greater than the ± 0.7 threshold are considered as ENSO events. El Niño years are shown in red while La Niña years are in light blue. Brown and blue bars show years when the La Niña and El Niño co-occur with the positive and negative IODZM events respectively.
Monthly mean NCEP/NCAR reanalysis data for the period 1948–2008 (Kalnay et al., 1996) are used to assess changes in the atmospheric circulation over the Indian Ocean region. Only the sea surface temperature (SST), wind stress and mean sea level pressure fields (MSLP) have been used herein because there is a general agreement that these variables are relatively reliable (Poccard et al., 2000; Nicholson and Webster, 2007). Since these variables are primary variables rather than those derived from other variables, (as is the case for the stream function, velocity potential, divergence and vorticity) they should be strongly influenced by observations, and hence more reliable (e.g. Kalnay et al., 1996).
Two datasets are used in the analysis of OND rainfall variability. Because it is gridded (2.5° resolution) and global in extent, the CAMSOPI (Climate Anomaly Monitoring System (‘CAMS’) and OLR Precipitation Index (‘OPI’)) dataset (Janowiak and Xie, 1999) is used to extract the dominant rainfall mode over the region using the principal component analysis (PCA) method. CAMSOPI is a monthly dataset covering the period 1979–2009 that merges rain gauge and satellite data from a number of sources. Since it is generally accepted that tropical temperate troughs (TTTs) are the dominant rainfall producing systems over southern Africa (e.g. Reason et al., 2006), this dataset is considered appropriate to represent the rainfall that emanates from the cloud bands associated with these systems. The fact that this high-resolution rainfall dataset is only available from 1979, mean that it is not long enough for the investigation of decadal climate variations.
However, observed monthly rainfall datasets with reasonably good spatial and temporal data coverage in the regions of dominant southern African rainfall variability do exist (Figure 3) over Zimbabwe and South Africa. These data enabled us to generate longer time series which can provide more information on interdecadal variations for the dominant rainfall variability. By restricting the missing data to only 10%, we were able to select a total of 17 stations, 10 from Zimbabwe, and 7 from eastern South Africa. These 17 stations were also deliberately chosen so as to coincide with the geographical location of the first principal component (PC1) loadings south of 15°S, with representative spatial distribution (Figure 3) close to and exceeding the 0.8 contour-loading values. Therefore, it is assumed that this data can describe the dominant variability of precipitation over southern Africa adequately. Gaps within these time series were closed using data from the nearest stations. As a result, we were able to construct continuous seasonally OND averaged time series with a common period of 1922–2008 (87 years). Although this method has some limitations, we considered it the best way to construct longer time series from the readily available data over the region.
Spatially averaged rainfall anomalies of the 17 stations (depicted in Figure 3) were used to calculate rainfall anomaly indices so as to better quantify the strength of the dominant rainfall variability. The mean and standard deviations were separately computed for each station and the respective new time series of normalized departures from the mean were formed. These standardized time series for each station were then spatially averaged in order to ensure that all locations contribute equally to the regional index as per Jones and Hulme (1996). At the same time, this procedure filters out noise produced by highly localized climate regimes common across the southern African region. This process also enabled the elimination of random components present in the data for individual stations and retained only commonalities over the selected region that can be considered as representative of larger scale conditions. In this way, the rainfall index for the OND dominant rainfall variability is constructed and designated as southern African OND rainfall index (SAOR) and its variability is shown in Figure 2.
In this study, Cramer's t statistic is used to compare the short-term persistent wet and dry periods of the southern African rainfall time series by applying 31-year running means. This statistic, is distributed as Student's t-statistic with (N-2) degrees of freedom, and can therefore be tested for significance (WMO, 1996) and herein, it compares the 31-year means for the sub-periods with the mean of the whole period. By so doing, we were able to isolate periods of surplus and deficit rainfall and also examine their significance. A similar approach was adopted by Kripalani and Kulkarni (2001), and Kruger (1999) on Indian monsoon and South African rainfall, respectively.
3.1. Extracting the dominant rainfall mode of variability over southern Africa
It is generally accepted that tropical temperate troughs (TTTs), which orient themselves from the northwest to the southeast across the region and link a tropical disturbance to a mid-latitude feature south of South Africa, are the dominant summer rainfall-producing systems over southern Africa. The variability in their preferred location between one summer and the next directly affects regional summer rainfall patterns (e.g. Washington and Todd, 1999; Tyson and Preston-Whyte, 2000; Reason et al., 2006). As a result, this mode of rainfall variability has long been considered as a feature of the mean summer climate of southern Africa.
Figure 3 presents the dominant rainfall pattern as created by the first principal component (PC1) spatial factor loadings. Here we note that unlike the spatial patterns for other PCs (not shown), it is only the spatial manifestation of PC1 that shows a single coherent system over southern Africa. This pattern, derived from highly homogeneous rainfall variability (loadings greater than 0.8) seems to reveal a rainfall system linking the central northwest through southern Africa to the southwest Indian Ocean (SWIO). Its configuration is therefore consistent with the synoptic scale TTT systems as defined in regional studies for the summer rainfall over much of southern Africa (e.g. Washington and Todd, 1999; Todd et al., 2004). It is for this reason that in the construction of the SAOR index we consider only station rainfall in the region coinciding with the most homogeneous rainfall that is extracted from PC1. As a result our approach pays more attention to the central eastern regions of southern Africa making it slightly different from techniques employed by most previous authors (e.g. Reason and Mulenga, 1999; Richard et al., 2001; Rouault and Richard, 2005).
3.2. Epochal relationship between SAOR and IODZM/ENSO
It is well known that significant long term changes in the distribution of the Indo-Pacific SSTs and tropical teleconnections occurred around 1976/1977 (Clark et al., 2000; Terray et al., 2005). In addition, Manatsa and Matarira (2009) noted that the two periods, 1960/1961 and 1997/1998 could be associated with major changes in the SSTs of the Indian Ocean, including the changing teleconnections of the IODZM with southern African rainfall variability. Figures 4 and 5 show the temporal evolution of the SST anomalies of the two poles of the IODZM and that of Niño 3.4 time series for the period 1948–2008 indicating how these two climate modes behaved during the identified two periods. A warming trend is apparent in these two climate modes though it is more significant with the IODZM. However, it is the eastern pole of the IODZM (Sumatra SST anomalies) which displays a more rapid rise with all the resulting four SST means of the epochs showing significant differences above the 95% confidence level using a two-tailed t-test. The western pole shows significant differences at this level only for the last three epochs (excluding the 1948–1960 period). It is interesting to note that Niño 3.4 only shows 1976 as a significant turning point as identified through the significance of the mean temperature differences. Of importance in these figures is the last epoch which shows 1997 as being a conspicuous turning point in all the time series. However, in the western pole of the IODZM, this last epoch identifies itself uniquely, with a falling trend that displays a confidence level of more than 99%.
Considering these consistent SST shifts in the Indian Ocean, it is appropriate to investigate if they also project themselves in the temporal association with regional rainfall. Thus, from a 31-year sliding correlation between IODZM and SAOR (inferred from Figure 6), we note that the association is not only unstable but is similarly punctuated by two major turning points; namely, 1961 and 1997. The relationship seems to be relatively strong during the earlier decades but collapses to insignificance after 1997. We could not easily identify the 1976/1977 climate shift impact in this analysis, thus suggesting that its possible contribution over the region is weak. Since this running correlation analysis hinted that the tropical Indian Ocean SST anomalies may exert a strong influence on the transitions of the IODZM and southern African rainfall, we therefore investigated the shifting nature of these correlations relative to the time evolution of the IODZM SST variance. Our examination of the link between the changing IODZM variance and the correlations is the key to demarcating the identified epochs. Most importantly we wanted to illustrate that the identified epochs in the previous research (Manatsa and Matarira, 2009) that was limited to covering Zimbabwe still holds more broadly within southern Africa.
Because of the apparent close relationship between correlation and covariance, the assumption is that with enhanced (suppressed) SST variability, one might expect enhanced (suppressed) teleconnections. Thus, 31-year sliding correlation values between the SAOR and IODZM are plotted on a scatter diagram against corresponding variance magnitudes of the IODZM index. In the scatter plot depicted in Figure 6(b), four clusters of homogeneous points can easily be identified. Coincidentally, these independently created four data clusters are non-random and fall almost perfectly into four time intervals that seem to represent three shifts to new and more stable states of inter-epoch relationships. Here, we note that the 1952–1961 and 1998–2008 clusters are distinctly separate from the other remaining two clusters. It is interesting to note that the 1952–1961 cluster and the 1977–1997 cluster assumed statistically significant negative linear associations (above 95% level) showing that during these epochs, periods of higher (suppressed) variability activated stronger (weaker) associations between the regional rainfall variability and the IODZM. However, this relationship is weakly identified in the linear relationship test of the last epoch and is visibly absent in the 1961–1976 epoch. Thus, during the period from 1952–2008, the epoch with strong linear relationship is preceded by an epoch with significantly suppressed association. However, the relative closeness (with some overlapping events) of the 1962–1976 cluster to the 1977–1997 cluster indicates similarities within these two epochs that possibly downplay the significance of the impact of the 1976/1977 climate shift as a turning point in this relationship. On the other hand, the relative remoteness of the clusters produced within the first and last epochs from the intervening epochs gives credibility to the consideration of 1961 and 1997 as representing major turning points.
This analysis has revealed that it is the 1998–2008 epoch that appears to be most different from the others. Though the pattern within the cluster is relatively organized, it does not show significant linearity relative to the preceding clusters. The mean correlation coefficients are smallest in magnitude relative to other epochs and the epoch has correlation coefficients that are insignificant above the 95% confidence level throughout. An F test showed that all the events within this last epoch have variances which are statistically different from those of the first epoch at 5% significance level, using a two-tailed t-test. Thus, although the IODZM is indeed in its maximum state of activity during the last epoch, there is virtually no corresponding strengthening in correlation with the rainfall; instead, weakening is observed. This result may indicate that the expected correlation covariance relationship in an assumed cause effect process may have been reduced substantially or ceased. The correlation coefficients that were initially highly dependent on the magnitude of IODZM index variance, with high variability corresponding to stronger connections, became independent of the IODZM index variance. It is apparent that this occurred immediately after the inclusion of the IODZM 1998 index value into the 31-year correlation segment. As a result, one may assume that the decoupling process between the IODZM and rainfall of southern Africa took place around this time. In the operational context, this observation suggests severe limitations to the usefulness of the IODZM index as a regional rainfall predictor after 1997, contrary to the recommendations of Manatsa et al. (2007, 2008).
To investigate how ENSO has behaved in its influence on the rainfall, the analysis is repeated but with Niño 3.4 index replacing the IODZM index. It is interesting to note that although 1961 and 1997 are quite vivid as turning points in the case of IODZM (Figure 6(b)), they are not diagnosed in the influence of ENSO on rainfall (Figure 6(a)). Instead, only the early 1970s break-up point, most probably related to the well known climate shift of the mid 1970s is distinguishable. This observation suggests that the decadal shifts of the 1960s and late 1990s remain significant only for the Indian Ocean regional influence but are not important in the ENSO-rainfall association.
Since the IODZM influence in the association between SAOR-ENSO is far greater than that of ENSO in the SAOR-IODZM relationship (Manatsa et al., 2008), partial correlations are used. In Figure 7, a scatter plot of the partial correlation between Niño 3.4 and SAOR when the IODZM signal is removed against Niño 3.4 variance is presented. Note that Figure 6(a) could lead to the false impression that ENSO has a relatively high and significant correlation with SAOR after 1973. Figure 7 corrects this misinformation by illustrating that these high correlations were mainly the result of the participation of IODZM index. We see that this significant correlation collapses to insignificance once the IODZM signal is statistically removed (e.g. Manatsa et al., 2008). Once again, the two clusters that are formed are similar to Figure 6(a) in that they seem to be organized but with no apparent linear links as depicted in Figure 6(b). A similar plot but using IODZM when ENSO is linearly removed does not show any significant deviation from Figure 6(b) and hence is not presented here. This result indicates that ENSO is less likely to be linearly influencing the OND rainfall variability of southern Africa. In fact, Fauchereau et al. (2009) noted that the apparent weakness of the ENSO regional impacts when linear metrics are used is due to the fact that the influence of ENSO is not all that linear over the sub-region.
To reduce the possibility of outliers biasing the results, we recalculated the IODZM/ENSO-SAOR simple and partial correlations using the Spearman correlation method. By its nature, this method is more resistant to outliers (Haylock et al., 2007). These new correlation series (not shown), closely resemble the series calculated using the Pearson correlation. In order to ascertain the robustness of this method, we also removed the years with the highest and lowest 5% of IODZM and SAOR values. Again, a similar epochal change in correlation was evident. Thus, we conclude that the shifts that we see in the link of the times series emanating from the IODZM variance and IODZM-SAOR correlations are not a product of several extreme years but are related more to slow changes in the climate. The post-1997 lack of a significant temporal relationship between IODZM variance and its correlation with SAOR is interesting. It automatically brings into perspective the question of cause and effect in relation to IODZM and southern African rainfall. Unlike in East Africa which is strongly influenced by the western pole of the IODZM, in southern Africa, the IODZM influence is more subtle. It is possible that the previously suggested IODZM influence was not direct, but through some other process. Unravelling this process is the prime focus of the following section.
3.3. Changes in the Indian Ocean SST anomalies during the epochs
The Indian Ocean is a significant source of moisture for the southern African region (D' Abreton and Tyson 1995; Reason et al., 2006) and hence a comprehensive assessment of the large-scale anomalous SST patterns in the SWIO is necessary in order to better explain the SST-regional rainfall connections. It is therefore expected that any changes in the Indian Ocean SST system might impact significantly on southern Africa rainfall. For example, the recently observed warming of the Indian Ocean (e.g. Terray et al., 2005) may be an important factor in explaining the changes in the noted IODZM-SAOR teleconnections via induced changes in regional atmospheric circulation. In this context, we investigated the behaviour of the Indian Ocean SST anomaly composites during the different identified epochs but only for the positive IODZM events. This bias in phase selection in the analysis is hinged on the mere observation that the influence of the IODZM events on southern African rainfall is not symmetric but is more pronounced in the positive phases (Manatsa et al., 2009). Therefore, it is through compositing the Indian Ocean SSTs associated with these positive IODZM events in different epochs that the sought differences in the changing influence of the IODZM during the record could be revealed.
To investigate the Indian Ocean SST changes associated with the noted shifts over the Indian Ocean, we considered all the positive IODZM years that occurred during 1948–2008 (1961, 1963, 1967, 1972, 1982, 1991, 1994, 1997, 2002, 2006, 2007, and 2008). We then assigned the corresponding events to the 3 epochs (i.e. 1948–1977, 1978–1997 and 1998–2008) and performed a composite study with respect to these three periods. Figure 8(a) to 8(c) show the Indian Ocean SST anomaly positive IODZM event composites for the three epochs. To examine whether there are any apparent changes in the Indian Ocean SSTs associated with the 1997 shift, Figure 8(d) plots the difference between the 1998–2008 and 1978–1997 epochs. Figure 8(a) and 8(b) have some similarities, with general warming over the western Indian Ocean and cooling over the Sumatra region accompanied by a rather strong cooling over the subtropical regions of the southwest Indian Ocean region. However, the composite of IODZM events for the 1998–2008 epoch shows that these most recent positive IODZM events occurred under basin wide warming except for the western IODZM pole where some relatively weak cool SST anomalies are still present. The difference in SST of this recent epoch from the previous epoch shows strong background warming in the southwest Indian Ocean, an area of high significance to southern African rainfall variability. Basing on previous atmospheric modelling studies (e.g. Reason and Mulenga, 1999; Reason, 2001), the SST changes in the southwest subtropical Indian Ocean are important as they are likely to be associated with substantial circulation alterations which may influence changes in local latent heat flux and low level moisture flux. However, the simultaneous warming and cooling over the western and eastern south Indian Ocean respectively in the most recent epoch relative to its predecessor may also imply an eastward shift of the south Indian Ocean anticyclone during the post-1997 positive IODZM events. Thus, assessing a possible zonal shift of the South Indian Ocean anticyclone is the focus of the next section.
3.4. The apparent eastward shift of the south Indian Ocean anticyclone
In this section, we consider whether the SST composite pattern for the last epoch is related to an eastward displacement of the semi-permanent South Indian Ocean anticyclone during positive IODZM events. To assess this possibility, we present the mean sea level pressure (MSLP) difference anomalies of the last two composites (Figure 9). We note that the results are consistent with what might be expected from quasigeostrophic theory (Gill, 1982) since low (high) pressure anomalies are approximately located near regions with warm (cold) SST anomalies south of Madagascar (west of Australia). This pattern suggests a general displacement of the South Indian Ocean anticyclone to the east during the positive phases of the IODZM.
The analysis is repeated to examine the averaged MSLP anomaly differences for the last two epochs, but inclusive of all the years in each period. A two-tailed t-test was used to determine the significance of the differences between the two epochs This test estimates the statistical significance above the 95% confidence level, at which the anomalies in the post 1997 period are distinguishable from the reference 1948 to the 1997 epoch at each grid point (Figure 10). The results suggest that the pre- and post 1997 changes in the anticyclone are statistically significant and are not just part of the positive IODZM events. In reality, repeating the analysis for the entire Southern Hemisphere (not shown) indicates that this shift is in fact due to a positive Southern Annular Mode pattern with its strongest anticyclonic anomaly in the midlatitudes located over the central South Indian Ocean. Thus, the notable MSLP changes that occurred in the southwest Indian Ocean (Figure 9) are not only restricted to the positive IODZM years but are consistent with the Southern Annular Mode tending to be more in its positive phase in recent decades (e.g. Marshall, 2003).
However, the region demarcated by p-values of 0.05 and above in the South Indian Ocean suggests that it is the western subtropical region that has the most obvious change. It is interesting to note that this region of significance also extends towards southeastern Africa from the subtropical region to coincide with the area previously identified as displaying the highest PC1 loadings (Figure 3). The statistically significant anomalous post 1997 behaviour of the South Indian Ocean anticyclone and its strong link to the southern African rainfall is further demonstrated by this observation. It is intriguing that this regime shift in surface pressure seems to be the most important event to occur in the Indian Ocean in the past century and that it has coincided with the noted post 1997 dramatic weakening of the correlation between the IODZM and the regional rainfall.
3.5. Changes in atmospheric anomalies over the South West Indian Ocean
The characteristic behaviour of the atmospheric anomalies in the subtropical south Indian Ocean is the focus of this section, particularly the region south of Madagascar. The identified western subtropical region of the south Indian Ocean herein has been found to be essential to the southern African rainfall variability in several previous studies (e.g. Goddard and Graham, 1999; Reason and Mulenga, 1999; Washington and Preston, 2006). Consistent with these studies we demarcate the area of anomalous warming with a box 25°S to 35°S and 35°E to 45°E (Figure 8(d)). Note that this box also covers the relatively stable southeasterly winds which normally blow towards southeast Africa. Anomaly time series of SSTs, pressure and u, v-wind stress components from which the corresponding indices are constructed, are extracted from this box. These indices are named southwest Indian Ocean anomaly SST index (SWIO_SST), southwest Indian Ocean anomaly MSLP index (SWIO_MSLP), southwest Indian Ocean zonal wind stress anomaly index (SWIO_U) and southwest Indian Ocean meridional wind stress anomaly index (SWIO_V), respectively. The temporal variability of these time series is then analysed. By showing how the magnitude of interdecadal variability of the various atmospheric parameters changes with time, we hope to provide an insight into the robustness of the 1997 shift outside the tropical Indian Ocean and hence the nature of the significant drop in the correlation between the IODZM and rainfall.
A closer scrutiny of the SWIO_SST (Figure 11) shows some interesting behaviour. The index had predominantly negative values in the earlier decades but reversed to positive in the late 1980s and then reverted to amplified anomalies after 1997. An 11-year running mean inserted in the graph illustrates this behaviour more clearly, especially the sustained positive gradient it assumed from the early 1970s. A further investigation was done using a Student's t-test for differences in the mean, to determine if in this boxed area, the SST increases in the last epoch depart significantly from the background variability in the SST data. Although there was a substantial statistically significant warming between the epochs 1961–1976 and 1977–1997 (p-value of 0.019), the most significant warming, as shown by a higher confidence level that nears 100%, is demonstrated in the epochal warming of the 1998–2008 era compared to its predecessor. Note that this established pattern in the differences of epochal means bears a qualitative resemblance to the western IODZM pole's SST anomaly epochal means shown in Figure 4(b). In both cases, the difference between the first and second epochal SST anomaly means is insignificant. However it strengthens to significance above the 95% level between the second and third epoch, but becomes even more pronounced between the third and fourth epoch. This result is indicative of temporal connections between SST anomaly related mechanisms over the IODZM and the SWIO.
If the South Indian Ocean anticyclone has indeed shifted eastwards as part of the suggested changes in the Southern Annular Mode and associated circulation, one would expect the demarcated south west subtropical region to have become less active, denoting a drop of variability in all atmospheric parameters including the SSTs. A visual inspection of Figure 12(a) is consistent with this hypothesis. We note that the period after the year 1995 was characterized by a statistically significant sharp fall in variance at 95% confidence level using an F-test. We later applied the Cramer's t-test to examine the stability of this record in terms of comparison between the overall mean of the entire record and the mean of the 31-year segment of the record. Figure 12(b) shows that the temporal manifestation of the Cramer's Statistic identifies the post 1997 period as a distinct epoch. To further support this finding we repeat the analysis method but substituting the SSTs with data sets of MSLP and u, v-wind stresses, respectively.
These further investigations using the pressure and wind stress variance changes in this southwest Indian Ocean subtropical region depict significant similarities in the pattern with the already analyzed SST variance. The 31-year SWIO_MSLP variances and their corresponding Cramer's t statistic demonstrate the period from 1992 to 1996 as the transitional period prior to the post-1997 era (Figure 13(a)). It is also interesting to note that even the temporal expression of the Student's t-test of significance for the Cramer's t-test values pinpoint 1997 as the period beyond which the Cramer's test values become significant at the 95% confidence level (Figure 13(b)). This pattern is consistent with slow changes in MSLP over the region, signifying some changes in the average characteristics of the South Indian Ocean anticyclone. Consistent with the 1997 drop in MSLP variance, the wind stress variance has also weakened. This fall is more pronounced in the meridional component of the wind stress, consistent with the post 1997 changes in the semi-permanent South Indian Ocean anticyclone relative to its pre-1997 behaviour. Thus, the strong drop in variance as well as the significant weakening of the surface winds to the south east of South Africa helps to account for the shifted IODZM–SAOR relationship.
However, it is interesting to note that these changes in the western subtropical South Indian Ocean did not profoundly affect the characteristics of southern African rainfall variability. By using the 31-year Cramer's t-statistic (Figure 15), noticeable changes are clearly shown in the temporal revelation of these statistic values, suggesting epochal changes in the OND rainfall totals. Although the post-1997 epoch can be inferred as a distinct epoch, it cannot be conclusively related to the 1997 shift as there is no corresponding shift in the rainfall but a manifestation of an interdecadal cycle. This result is not surprising as much of southern Africa has long been known to experience significant decadal to interdecadal variability in rainfall (e.g. Tyson and Preston-Whyte, 2000; Reason and Rouault, 2002; Allan et al., 2003). A test for significance shows that the Cramer's t-test values are not significant even at a lower 80% confidence level. That could be the reason why the rainfall epochs shown in Figure 14 were not detected as distinct using the Student's t-test for the differences in the epochal means, nor was a significant trend observed in the whole time series using the Mann-Kendal test analysis. Thus, the post-1997 substantial loss of statistical significance in the relationship between the IODZM and regional rainfall does not seem to be associated with any major statistical changes in the SAOR time series themselves.
3.6. Connection of IODZM to the subtropical southwest Indian Ocean region
It has already been observed that the temporal manifestation of the IODZM is qualitatively similar to that of the subtropical southwest Indian Ocean region. This observation indicates a possible link between these two sub-regions. To see if this is true, we applied a simple cross-correlation analysis for moving 31-year segments, as well as for the entire period. Table I shows the results of the cross-correlation among the indices used in this study. We note that the correlation values for IODZM with SWIO_MSLP and the SWIO_SST index are statistically significant at the 90% (0.31) and 99% (0.46) confidence levels, respectively, when the whole 1948–2008 period is considered. This result implies that for the OND season the IODZM could be connected to variability in the subtropical southwest Indian Ocean MSLP, SST and meridional wind stress time series. However, the correlation value of the IODZM with the zonal wind stress index indicates otherwise, as the association is statistically insignificant with a coefficient value of 0.18. As with previous research (e.g. Matarira, 1990), correlations involving entire periods could be misleading as they obscure possible non-stationary relationships. It is interesting to note from Table I that the SWIO_MSLP is significantly correlated (above the 99% confidence level) to the meridional component of the wind stress whilst insignificantly linked to the zonal wind stress, with less than twice the value of the latter. At the same time, the SAOR is significantly inversely linked at 90% confidence level to the meridional wind stress component and weakly negatively linked to the surface pressure. The connection with the zonal wind stress component seem to be negligible (r = 0.08).This observation is of regional importance in the sense that it reflects the paramount role played by the meridional wind component in modulating the regional rainfall. Thus, in the southeasterly winds that have been noted to significantly influence the regional rainfall, it is most probably the meridional wind element that is most active. On the other hand, the strength of this meridional wind is controlled by the steepness of the land-ocean meridional gradient over the region which subsequently entails the level of stability of the whole south easterly airmass affecting southern Africa.
Table I. Cross-correlation coefficient results among the six indices for the period, 1948–2008. Bold and italics stand for significance above 99 and 90% confidence levels, respectively
SWIO_ SIO_ SST
Figure 15 illustrates the spatial correlation patterns of the IODZMI with the meridional wind stress component in the Indian Ocean basin for the whole period of analysis from 1958 to 2007 (a), the pre-1997 period (b) and the post-1997 period (c). We only show the meridional wind stress component correlations as it is the component that identifies the subtropical South Indian Ocean as an important link region with the IODZM and not the zonal component. Once again, the importance of the meridional wind component in the IODZM impact on southern Africa is demonstrated. However, of importance here is the realization that considering the whole period in investigating the associations can mask some of the important epochal changes within the period. The correlation patterns in the left and middle panels are qualitatively similar giving a false impression that the IODZM spatial correlation characteristics with the meridional wind stress element have not changed significantly for the two periods. But on isolating the correlation pattern of the post-1997 period we note that this pattern has been altered substantially. We see that the area covered by correlation values that are significant at p values less than 10% have generally shrank considerably over the whole region. But of relevance to this study is the complete disappearance of the region to the south east of South Africa that had prior significant linkages with the IODZM. This means that after 1997, the IODZM could have been disconnected linearly from the subtropical southwest Indian Ocean surface meridional wind.
Further investigations using the running correlation analysis, whose results are shown in Figure 16(a) and (b), points towards the prominent occurrence of this 1997 shift and shows it as not only an oceanic but also an atmospheric event. The analysis reveals conspicuous epochal correlation variations which once again identify the period from 1997 as an epoch with distinctly different properties. This result is also true for the temporal associations of the wind stress components with the remote IODZM and the SWIO_MSLP anomalies over the south west subtropical Indian Ocean. These observations provide further evidence for a shift to a new quasi-stable state since the change in the wind stress u, v components, the underlying MSLP and the IODZM index, occurred around 1997. Thus, the 1997 ocean and atmospheric changes that occurred over the Indian Ocean seem to be of basin-wide scale in character, occurring both within and outside the tropics, at least covering the area between 10°N and 50°S. Thus, we suspect that the observed significant post 1997 changes that occurred in the subtropical SST anomalies (Figure 4(b)) may have triggered simultaneous major changes to the regional circulation anomalies resulting in the post-1997 weak correlation values that we note here. The observed impact of 1997 on the variance has proved to be sudden and profound, indicating that it is inevitable that a shift in temporal teleconnection patterns involving the IODZM and the basic atmospheric parameters over the subtropical southwest Indian Ocean has occurred.
3.7. SAOR relationship with the southwest Indian Ocean MSLP and SST anomalies
The use of the 31-year running correlations indicates that a sharp drop to insignificance in the association between the SAOR and MSLP or SST indices occurs as soon as the respective index values for the year beyond 1997 are included in the 31-year correlation segment. Thus, Figure 17 shows the IODZM-SAOR and IODZM–SWIO_SST temporal connections using 31-year sliding correlations. The results reveal that both relationships experienced similar interdecadal variations, with relatively strong links during the pre-1997 decades that weaken abruptly immediately after. Thus, these two indices might have been useful predictors for OND rainfall before 1997. However, after this year, their potential for this purpose is eroded completely. The fact that these major correlation transitions to insignificance occurred almost simultaneously in both relationships implies that both the IODZM–SAOR and the IODZM- SWIO_ SST relationships could be associated with a common interdecadal signal.
The event centred around 1997 could be one of the most important cases of abrupt climate shift to occur in the Indian Ocean climate system in the 20th century. This occurrence is of great importance for southeastern Africa since the south easterly winds from this region are an important determinant of summer rainfall over the region (Tyson and Preston-Whyte, 2000; Reason et al., 2006). But since we noted earlier that it is the meridional wind component in the south easterly winds that play a bigger role in influencing the regional rainfall, attention is shifted to this wind. Its averaged seasonal strength is strongly influenced by the surface pressure characteristics of the semi-permanent South Indian Ocean anticyclone by way of altering the meridional ocean-land pressure gradient due to the anticyclone's assumed quasi position relative to the South African coast. When the gradient is strong (anticyclone close to the land) the wind is also strong and advects colder meridional air. But because these pole-ward winds are generally stable with moisture confined mostly to their lower levels, convergence with relatively unstable tropically based air masses from the north is displaced further north out of the sub-region resulting in rainfall deficits (Tyson and Preston-Whyte, 2000). On the other hand, a slackened meridional pressure gradient and the developed relatively low pressure resulting from the displaced or weak anticyclone encourage the formation of the TTTs and hence surplus rainfall. That is why we see in Table I that the link between SAOR and the meridional (v) component of the surface winds is inverse (stronger surface v-winds associated with rainfall deficits and the reverse) and significant at the 90% confidence level, but with virtually no link with the zonal wind stress factor.
We have noted that in the pre-1997 era, the IODZM was positively related to the western subtropical South Indian Ocean MSLP, SSTs, zonal and meridional winds, but inversely related to the regional rainfall. These relationships mean that the positive IODZM events were mostly associated with high pressure anomalies over and near the south east South African coast which intensified the meridional component of the south easterly winds and hence increasing the stability of this airmass. Ultimately this process could have significantly reduced the chance of the development of TTT rainfall producing weather systems over the region and increased the likelihood of dry conditions. This explains why Manatsa et al. (2009) noted a very strong connection of regional droughts and the positive IODZM but with no obvious link with the negative IODZM.
However, this process changed profoundly for the post-1997 positive IODZM events that occurred when the South Indian Ocean anticyclone has shifted further to the east. Under these conditions, the circulation east of South Africa became more cyclonic, thereby favouring the development of TTTs and, hence, increased rainfall. The variance of the SWIO_MSLP and the related SWIO_SSTs and wind stress components were likewise substantially reduced in the process. Consequently, the post-1997 positive IODZM events did not act to enhance the likelihood of droughts over southeastern Africa through enhancing the meridional component of the southeasterly winds and suppressing the TTT formation, as was the case for those events before 1997. At the same time, the variability of the IODZM showed that it is independent of the 1997 eastward shift of the South Indian Ocean anticyclone and hence did not show a corresponding post 1997 reduction in variance as was the case with other climatic variables over the west subtropical south Indian Ocean. Thus, the weakening in anticyclonic conditions over the western subtropical South Indian Ocean does not appear to be in harmony with the variability of the rainfall over adjacent southeastern Africa since there is a collapse of significance in the correlations between the rainfall and the climate parameters over this ocean region for the period after 1997. It then becomes apparent that any IODZM influence on the rainfall variability over southern Africa is not direct but occurred through its close association with the zonal location of the south Indian Ocean anticyclone. Once the anticyclonic conditions to the south east of South Africa were weakened by the eastward shift of the south Indian Ocean anticyclone, the tendency of positive IODZM events to be in phase with regional droughts of in the OND season, was either greatly reduced or eliminated.
4. Summary and conclusions
Changes in variability characteristics are used in this study to illustrate that the regime transitions in the Indian Ocean are found around 1960/1961 and 1997/1998, but with the intervening well known 1976/1977 climate shift being of lesser importance. However, it appears that the transitions in the epochal connections are externally induced. The pre-1997 period is characterized by significant negative SST anomalies over the western subtropical South Indian Ocean, at least during the positive IODZM years, while the post-1997 era shows a reversed SST anomaly distribution. These changes in SST anomalies can be inferred to indicate the weakening of anticyclonic conditions east of South Africa after 1997, hence implying an eastward shift of the semi-permanent South Indian Ocean anticyclone. It is therefore the zonal movement of this anticyclone during the positive IODZM events that could have played a pivotal role in triggering the observed transitions of the link between the IODZM and SAOR. Thus, the IODZM was associated with southern African rainfall variability in the OND season through its connection with the variability of the South Indian Ocean subtropical anticyclone when the latter was close to the southern African coast. It appears that IODZM was decoupled from the rainfall as soon as the eastward movement of the South Indian Ocean anticyclone became too great. This finding suggests that changes in the state of the background Indian Ocean climate have an influence on the predictive relationships between IODZM and southern African rainfall.
The regional climate of the subtropical southwest Indian Ocean may have a tendency to exist in quasi-stable states for prolonged epochs as demonstrated by the epoch before and after 1997. If this change around 1997 persists and is fairly stable, then the connections between the positive IODZM and OND rainfall over southeastern Africa should remain weak. The next decade or two will help in assessing whether this occurs or not. However, within the confines of the data record length, 1997/1998 appears to stand out as a climate regime shift since it represents the single largest phase transition in the interdecadal oscillations during the study period.
Consequently the results of the current work suggest that seasonal predictions of OND rainfall from Indian Ocean SST patterns may be subjected to interdecadal variations in the forecasting skill. Although this skill variation may be problematic, a greater understanding of the physical mechanisms that drive the observed changing climate states enables appropriate evaluation of the stability of predictive rainfall relationships. In this way, it should be possible to identify and establish a predictive scheme that is more stable and useful within the context of a shifting background climate, be it through natural or anthropogenic causes.
The authors wish to thank the free provision of NCEP/NCAR data and the construction of diagrams using Climate explorer facilities. Provision of facilities and partial financing from Bindura and Cape Town Universities is greatly appreciated. NRF funding that assisted in this writing is acknowledged. Comments from the two reviewers are also greatly appreciated as they shaped the final state of the paper.