The influence of ENSO on winter rainfall in South Africa


  • N. Philippon,

    Corresponding author
    1. Centre de Recherches de Climatologie, UMR5210 CNRS, Université de Bourgogne, Bâtiment Sciences Gabriel, 6 blvd Gabriel, 21000 Dijon, France
    • Centre de Recherches de Climatologie, UMR5210 CNRS, Université de Bourgogne, Bâtiment Sciences Gabriel, 6 blvd Gabriel, 21000 Dijon, France.
    Search for more papers by this author
  • M. Rouault,

    1. Department of Oceanography, Mare Institute, University of Cape Town, Rondebosch 7701, South Africa
    2. Nansen-Tutu Center for Environmental Research, University of Cape Town, Rondebosch 7701, South Africa
    Search for more papers by this author
  • Y. Richard,

    1. Centre de Recherches de Climatologie, UMR5210 CNRS, Université de Bourgogne, Bâtiment Sciences Gabriel, 6 blvd Gabriel, 21000 Dijon, France
    Search for more papers by this author
  • A. Favre

    1. Climate Systems Analysis Group, ENGEO department, University of Cape Town, Private Bag X3, Rondebosch 7701, South Africa
    Search for more papers by this author


Whereas the impact of ENSO on the African summer rainfall regions is largely documented and still regularly investigated, little is known about its impact on the winter rainfall regions located at the southwestern and northwestern tips of Africa. Yet, these regions are densely inhabited and are net exporters of high-quality agricultural products. Here we analyze the relationship between El Niño Southern Oscillation (ENSO) and South Africa austral winter rainfall using a 682 raingauges daily rainfall database documenting the period 1950–1999. The May, June and July (MJJ) seasonal rainfall amount shows a positive correlation with the Niño3.4 index that becomes significant since the so-called 1976/1977 climate regime shift. Wet spells properties (length, frequency and intensity) at the raingauge scale indicate that high (low) MJJ seasonal rainfall amounts recorded during El Niño (La Niña) events are the result of longer (shorter) wet spells in the Cape Town area and more (less) frequent wet spells north of 33°S. Wet spells with daily rainfall amounts ranging between 10 and 50 mm are also more (less) frequent. Atmospheric dynamics fields during wet spells feature lower (higher) pressure and northwesterly (southerly) wind anomalies in the troposphere over the region. This suggests that rain-bearing systems are deeper (thinner) and larger (smaller) in extent, and located farther north (south) during El Niño (La Niña) events. Copyright © 2011 Royal Meteorological Society

1. Introduction

Southern Africa receives most of its rainfall in austral summer except for a region in the southwest that experiences austral winter rainfall. Rainfall maxima are recorded from May to August (Rouault and Richard, 2003) when the track of the temperate weather systems (i.e. extratropical cyclones, cold fronts and cutoff lows) is shifted northward. That southwestern region which encompasses part of the Western and Northern Cape Provinces is bordered in the east by the cold Benguela upwelling system, and to the south, at a distance, by the warm Agulhas Current. Orography plays an important role. In particular, the Cape Folded Mountains which stretch northward to the west and east–west to the south (red dashes, Figure 1) favour high (low) rainfall on their seaward (landward) side. The agricultural sector is critical for the economy of this region (e.g. it accounts for 60% of the Western Cape regional exports). A variety of export-grade fruits (apples, table grapes, oranges …) and wines are produced with grapes and deciduous fruits being mainly cultivated under irrigated conditions.

Figure 1.

Elevation map of the study region based on SRTMv2 data (Farr et al., 2007). The white dashed lines indicate the Cape Folded Mountains (∼2000 m) which stretch north–south to the west and east–west to the south. (1) Hottentots Holland, (2) Drakenstein and (3) Cederberg mountains within the Cape Folded Mountains

While summer rainfall in southern Africa is known to be influenced by the ENSO (dryer than normal conditions during El Niño and wetter than normal conditions during La Niña events, Lindesay et al., 1986; Richard et al., 2000; Mason, 2001; Reason and Rouault, 2002; Misra, 2003; Kane, 2009; among others), the winter rainfall interannual variability was related to the Antarctic Annular Oscillation (Reason and Rouault, 2005), Antarctic sea ice extend (Reason et al., 2002; Blamey and Reason, 2007) and South Atlantic sea surface temperature (Reason et al., 2002; Reason and Jagadheesha, 2005). No relation was found with ENSO (Blamey and Reason, 2007; Reason and Rouault, 2005). However, those studies considered an extended winter season from May to September, and long study periods (1900–2000, or 1950–2000) during which the rainfall-ENSO teleconnections have changed (Janicot et al., 1996; Richard et al., 2000; among others). Recently, Rouault et al. (2010) found an impact of the ENSO on the Western Cape summer climate and adjacent sea surface temperature where the prevailing southeasterly winds drive a coastal upwelling. Wind speed during El Nino (La Nina) events is weaker (stronger) than normal in the region leading to changes in sea surface temperature.

Given the importance of the agricultural sector for the economy of the region, the high rate of irrigation farming practised, and the predictability conveyed by ENSO, we reassess the teleconnection between ENSO and rainfall for the winter rainfall region of South Africa. Using a high-density daily rainfall dataset and reanalyzed climate data, we show that the seasonal rainfall amounts and several wet spells properties and associated atmospheric dynamics are significantly modulated during El Nino and La Nina events since the 1980s during austral winter.

The paper is divided into 4 sections. Section 2 presents data used. Section 3 is devoted to results. They are shared between four subsections: the first one deals with the ENSO signal in rainfall seasonal amount; the second explores the shifts in wet spells characteristics (i.e. number of wet spells, their length and intensity) during ENSO; the third presents anomalies in atmospheric dynamics fields related to wet spells during positive and negative phases of ENSO; in the fourth subsection the potential for seasonal rainfall forecasting is discussed based on lead-lag correlation analyses. A discussion closes the paper in Section 4.

2. Data

2.1. Rainfall

The daily rainfall dataset documenting the southwestern part of the Republic of South Africa were obtained from the Water Research Commission rain-gauge database compiled by Lynch (2003). First, we extracted 1187 stations located in the domain 35°–29°S/17°–24°E and presenting no missing values for the 1950–1999 period. Then, we applied a cluster analysis on the 365 daily means to identify those stations recording a rainfall peak in austral winter. The initial number of Classes was set up to 20 and after a close look at the cluster tree we retained only 10 Classes. Figure 2(a) and (b) presents the results of the cluster analysis with the 10 retained Classes. Panel (a) displays the locations of the stations and their Classes, while panel (b) presents the annual cycle for each Class (filtered with a 30-day running mean). Stations belonging to Classes 1–6 record most of their rainfall from April to September, with a peak in June. They are all located west of 21°E. The amplitude of the annual cycle differentiates the 6 Classes. Class 1 includes most of the stations and has a maximum of around 1 mm/day in June. The wettest Classes (2–5, 3–8 mm/day in MJJ) are in the vicinity of Cape Town and on the seaward side of the Drakenstein and Hottentots Holland mountain ranges (Figure 1). Classes 7 and 8 are located on the Garden Route, an area receiving rainfall all year around with slightly higher amounts in spring and fall. Lastly, Classes 9 and 10 which are located on the high plate of the Karoo or on the landward side of the Cape Folded Mountains (Figure 1), exhibit a summer rainfall maximum (peak in February). Given that stations falling in Classes 7 and 8 receive appreciable rainfall in winter related to the eastward displacement of low-pressure systems that also affect stations belonging to Classes 1–6, we are keeping Classes 7 and 8 in our analysis. However, we have excluded from the analyses stations belonging to Classes 9 and 10 only. This leaves us with a total of 682 stations (belonging to Classes 1 to 8). This clustering performed at the seasonal time step agrees with Mason (1998) who used 430 rain gauges across South Africa to delineate regions with homogeneous interannual variability over the period 1951–1995. In particular, our Classes 1–6 and 7–8 are in a good match with the regions referred to as G and H, respectively, in Mason (1998, in figure 1 herein).

Figure 2.

(a) Location and cluster number of the 1187 stations extracted from the Lynch database over the domain 29–35°S/17–24°W, and the period 1950–1999; (b) mean rainfall annual cycles of the 10 clusters filtered with a 30-day moving window

To complement the Lynch database which has not been updated after 1999, we used monthly gridded precipitation data provided by the Global Precipitation Climatology Centre (GPCC). We selected the 0.5° resolution ‘reanalyzed’ dataset (V.4) which offers an optimized spatial coverage for the period 1901–2007 (Rudolf and Schneider, 2005). On average, about 250 stations document the grid points over our region of interest. However, during the years 2006–2007, the number of stations drops dramatically under 50 (not shown). As a consequence, we restrict our analyses to 2005. In addition, to be consistent with the Lynch database, we considerer only the grid points for which station records are available.

2.2. Sea surface temperatures

We used the monthly Niño3.4 index (5°S–5°N/120°–170°W) over the period 1950–2005 based on Hadley Centre SST dataset HadISST available at the Royal Netherlands Institute of Meteorology (KNMI) climate explorer (Van Oldenborgh and Burgers, 2005). Monthly values were then averaged per trimester: March–May (MAM), May–July (MJJ) and July–September (JAS). El Niño and La Niña events are defined as the ones when the anomaly of the trimester considered is respectively above 0.5 std and below − 0.5 std. However, given the life cycle of ENSO, the March–July months correspond to either the onset phase or the decay phase (Larkin and Harisson, 2002). This distinction between the two phases must be kept in mind when we assess the relationship with rainfall. Indeed, we can expect the relationship to be modulated according to the phase of ENSO.

2.3. Atmospheric dynamics

We used the NCEP-DOE AMIP-II (NCEP2) reanalyses (Kanamitsu et al., 2002) to infer daily atmospheric dynamics over the domain 0–50°S/10°W–80°E during MJJ. Five parameters—meridional (V) and zonal (U) wind, geopotential height (Z), relative humidity (H) and temperature T—at two levels (850 and 500 hPa) were considered. Note that for the 850 hPa level, relative humidity was converted into specific humidity and then multiplied by U and V to obtain humidity flux.

3. Analyses

3.1. ENSO signal in seasonal rainfall amounts (synchronous relationships)

The seasonal rainfall amounts—Niño3.4 index relationships are presented for the beginning, core and end of the winter rainy season, i.e. the following three overlapping trimesters MAM, MJJ and JAS (note that analyses for AMJ and JJA were performed as well and lead to results very close to MJJ which is the trimester recording the highest rainfall total). The relationships with the Niño3.4 index are considered synchronously, e.g. MJJ rainfalls are correlated with MJJ Niño3.4.

To highlight a possible decadal variability in the teleconnection with ENSO (Janicot et al., 1996; Richard et al., 2000), we perform a Principal Component Analysis (PCA) on the seasonal rainfall amounts of the 682 stations for each trimester and each 21-year running sub-period of the 1950–1999 period. Then, we correlate the first three PCs (carrying, on average, ∼52, 16 and 6% of the total variance, respectively) to the Niño3.4 index. We perform the same procedure using the GPCC gridded rainfall over the period 1950–2005. Figure 3 presents the time evolution from 1950 to 2005 of the correlations for the 1st PC only which is the only one to show significant and robust correlations with the Niño3.4 index (note that over their 1950–1999 common period, the correlation between Lynch and GPCC 1st PC equals 0.94). The correlations significance (and in the remainder of the paper) is tested following the ‘random-phase’ test approach by Ebisuzaki (1997). For each pair of time series to be correlated, 100 random time series having the same power spectrum as the original one, but random phases, are correlated, and the linear correlations are ordered in ascending order. The upper 5% tail gives the one-tailed 5% level of significance with the alternative hypothesis that the correlation is positive. Figure 3 shows that over the last 30 years, there is an increasing correlation between MJJ rainfall and the Niño3.4 index, while no robust and significant correlation is found for MAM and JAS. Richard et al. (2000) noticed a very similar strengthening of the ENSO association with the summer rainfall over the north and east parts of the RSA since the mid 1970s. They emphasize the role of the 1977 shift in southern oceans temperature from cold to warm conditions.

Figure 3.

The 21-year moving window correlations between the Niño3.4 index and the rainfall amount 1st PC on the period 1950–2005 from Lynch (1950–1999, line) then GPCC (2000–2005, dots) for the three trimesters MAM, MJJ and JAS. Niño3.4 and rainfall are considered synchronously. The black circles denote values significant at the 95% level according to Ebisuzaki random phase test. The x-axis provides the central year of the 21-year period. This figure is available in colour online at

In the following, given the strong decadal component of the teleconnection, we focus on the 21-year period 1979–1999 for which daily rainfall from Lynch database and NCEP2 reanalyses are available and when the correlation with Niño3.4 is above 0.5.

First, the scatter-plot between Niño3.4 and the MJJ rainfall computed from GPCC over the period 1979–2005 (but standardized over 1979–1999 to match the Lynch database period) is given in Figure 4. El Niño and La Niña years and their phase (onset, peak, decay) as defined by CPC ( over the period 1970–2000 are represented with circles of different colours. The correlation equals 0.51. It is obvious that not all the El Niño (La Niña) events are associated with strong positive (negative) rainfall anomalies. For instance, during the 1982 and 1987 warm events (the 1985 cold event), rainfall anomalies are close to zero or are of a reverse sign. Moreover, the intensity of the rainfall anomalies seems completely independent of the phase of the events (onset, peak or decay) whereas, one could expect the decay phases to be associated with larger anomalies than the onset phases. Lastly, there is no relation with the type of event, canonical or Modoki (Ashok et al., 2007).

Figure 4.

Scatter-plot between Niño3.4 (x-axis) and the rainfall amount 1st PC (y-axis) from GPCC over the period 1979–2005 and the trimester MJJ. Data are standardized over the period 1979–1999. Empty, dark blue (red) and light blue (red) circles (triangles) denote onset, mature and decay phases of La Niña (El Niño) events as defined by CPC over the period 1970–2000. This figure is available in colour online at

Further, it is interesting to note that the 1997/1998 El Niño event that had a lesser impact than the 1982/1983 event on the summer rainfall of Southern Africa (Rouault and Richard, 2005; Lyon and Mason, 2007, 2009), is associated with strong winter rainfall anomalies over the Western Cape. Conversely, the 1991–1992 event which had a very strong impact on the summer rainfall of Southern Africa (Rouault and Richard, 2005), is associated with weak winter rainfall anomalies over the western region.

Figure 5 displays the spatial pattern of the correlations between the Niño3.4 index and seasonal rainfall amount for each of the 682 stations on the period 1979–1999. The three trimesters are considered to account for spatial shifts that could have been masked by the PC. In addition to the local significance testing, the field significance has been tested as well. Indeed, given the large number of stations used and their close proximity, the substantial station-to-station correlation can yield to spatial coherent areas of chance sample correlation. The field significance has been assessed using the False Discovery Rate (FDR) approach initially proposed by Benjamini and Hochberg (1995). This approach has been proved to be powerful in evaluating field statistical significance when many hypothesis tests are performed simultaneously with climatological data (Ventura et al., 2004), and to be insensitive to the non-independence of the local test results (Wilks, 2006). Following Brown et al. (2010), we have computed for each correlation field the proportion of stations having a p-value < = 0.1 and have calculated the q-value that is found for p = 0.1. This gives the expected proportion of false positive correlations incurred for correlations assumed significant. When all the correlations are false positive, the field is assumed to be not significant.

Figure 5.

Correlation maps of the Niño3.4 index and synchronous seasonal rainfall amount for the 682 stations and the 3 trimesters MAM, MJJ and JAS. The red (blue) shades denote positive (negative) correlations. Circles with blue and red contours denote values significant at the 95% (90%) level according to Ebisuzaki random phase test. The percentage of significant correlations at α = 0.1 and the False Discovery Rate (FDR) are also provided when the field is significant

In agreement with the preceding results, a majority of stations display a weak positive correlation with Niño3.4 in MAM and logically the correlation field is not significant. In MJJ, much larger (and significant) positive correlations are observed mainly for the stations located north of 34°S (i.e. those significantly described by the 1st PC). Few stations located on the south coast (along the Garden Route) are significantly correlated with the Niño3.4 index as well. It makes a total of 40% stations having a correlation significant at α = 0.1 with 4% only being false positive according to the FDR method. The contrasted behaviour of the south coast stations is obvious in JAS. While the correlations with Niño3.4 switch from positive to negative for the stations west of 20°E, they are still positive (and even larger than in MJJ) for the stations of the Garden Route. However, most of these correlations are not significant at α = 0.1 and the correlation field is not significant.

3.2. ENSO signal in the rainy season components properties (synchronous relationships)

In this subsection, the focus is on the MJJ season, core of the winter rainfall season. The daily resolution of the Lynch dataset allows us to compute some intra-seasonal properties and to explore the impact of ENSO therein. Indeed, it is interesting to know if the anomalies observed in seasonal rainfall amount during El Niño and La Niña events come from shifts in the frequency of rainy days and/or in their intensity. Ropelewski and Bell (2008) analyzed such shifts in South American rainfall. They noticed that whereas no clear signal could be found in seasonal rainfall amounts, daily amounts of 20 mm/day and up were significantly modulated during El Niño and La Niña events. For the South African summer rainfall, Crétat et al. (2010) demonstrate that the correlation between the Multivariate ENSO Index and DJF frequencies of dry days reaches 0.75 for the period 1970–1999, while the seasonal correlation between rainfall total anomalies and ENSO is of the order of 0.5. In this study, three characteristics of the rainy season are examined: the number of wet spells (NWS), their length (LWS), and their intensity (IWS) which is computed for each wet spell following:

equation image(1)

where AWS is the rainfall amount recorded during the wet spell. A wet spell is defined as a series of rainy days (> = 1 mm), and the shortest wet spells are 1-day long. Note that the seasonal amount S can be retrieved from

equation image(2)

Similar to Figure 5, Figure 6 presents correlation between the Niño3.4 Index and NWS, LWS and IWS at the station scale during MJJ. LWS and IWS display broad patterns of positive correlations suggesting that during El Niño events, wet spells are longer and bring higher rainfall amounts than during La Niña events. LWS is significantly modulated in Cape Town area, east of 20°E. NWS pictures a contrasted pattern with weak negative correlations in Cape Town area and along the Garden Route, and high positive ones stretching from Lamberts Bay to Namaqualand northwards. Only the correlation field for IWS is significant. This is not surprising given that wet spells intensity spatial correlation is usually less than the wet spells length or wet spells number spatial correlation. 26% of stations are significant at α = 0.1, with 19% of them being false positive. To sum up, the high seasonal rainfall amounts recorded during El Niño events are the result of longer wet spells bringing more rains over Cape Town area and more frequent wet spells bringing more rains north of 33°S. This suggests that El Niño events in austral winter could be associated with mid-latitude storm tracks shifted to the north and/or with larger and deeper systems. Reason et al. (2002) found that those conditions occurred during wetter than normal winter in the Western Cape. One way to verify the second point, is to compute the conditional probability of occurrence of a wet day (>1 mm) around a wet station as a function of the distance for El Niño events (Niño3.4 > 0.5std: 1982, 1983, 1987, 1991, 1992, 1993, 1997), La Niña events (Niño3.4 < − 0.5std: 1981, 1984, 1985, 1988, 1989, 1998, 1999) and neutral events (1979, 1980, 1986, 1990, 1994, 1995, 1996). Results for the three samples (of 7 events each) are presented in Figure 7. As expected, the probability decreases with distance for the three samples (the farther away from the wet station of reference, the weaker the probability of recording rainfall). However probabilities are systematically higher (lower) for El Niño (La Niña) events and departures become significant at 95% (according to a Student t-test) when the distance from the reference station ranges between 480 and 720 km. This result indicates that the rainy systems affecting our study region in MJJ have a larger extent during El Niño events than during La Niña events.

Figure 6.

Same as Figure 5 but for the MJJ wet spells properties, i.e. the frequency (NWS), the length (LWS), and the intensity (IWS)

Figure 7.

Probability of occurrence of a wet day (> = 1 mm) around a wet station (average of the 682 stations) according to the distance (in km) for El Niño (red thin dashed line), neutral (full black line) and La Niña (blue thick dotted line) events. The circles denote differences significant at the 95% level according to the Student's t-test. This figure is available in colour online at

To further document the impact of ENSO on the three intra-seasonal characteristics of daily rainfall and following Ropelewski and Bell (2008), we test the significance of the shifts induced in their distributions using a Kolmogorov-Smirnov (KS) test (Smirnov, 1948). This test aims at determining if two independent samples are drawn from the same continuous population or not. To account for differences in El Niño and La Niña events which do not have symmetrical impacts, the KS statistics are computed on the three samples considered two by two: El Niño versus neutral events, La Niña versus neutral events, and El Niño versus La Niña events. The mean distribution of the three components along with the differences between samples are presented in Figure 8 for all stations. All results are significant at the 95% level and are strengthened when the 1982 and 1987 events are excluded from the El Niño sample. On average (Figure 8, black thick line), during MJJ, the western region of South Africa typically experiences 11–12 wet spells lasting 1–2 days and bringing about 5 mm/day, a pattern significantly modified during El Niño and La Niña events. For instance, during El Niño events, the number of wet spells (NWS) tends to be shifted toward 15 per season and per station. This increase in NWS can be attributed to the increase of NWS at the northernmost stations (also in Figure 6 (a)), and reflects larger or rain-bearing systems reaching farther north rather than more frequent systems. In addition, the 10–50 mm/day wet spells (IWS) are more frequent. La Niña events are associated with a reverse pattern for IWS: less intense wet spells are more frequent with a threshold set around ∼6 mm/day (against 10 mm/day for El Niño years). Actually, composite maps of the variation (in % of the normal) of the wet spells intensity for El Niño and La Niña years (not shown) indicates that the signal comes mainly from La Niña, with most of the stations recording wet spells 20–40% less intense during La Niña years than during neutral years. Moreover, La Niña events also seem to be associated with a higher number of wet spells (NWS) as compared to neutral events but this could be an artefact induced by the higher number of one-day wet spells (LWS). Similarly, the composite maps of the variation in % of normal of the number of wet spells (not shown) for El Niño and La Niña years show that NWS increased (decreased) by 20% for stations north of 32°S.

Figure 8.

Thick black line: histograms of the number (NWS), length (LWS), and intensity (IWS) of wet spells for the 682 stations in MJJ and all the years. Bars: histogram of the difference between El Niño and La Niña events. Thin black line: histogram of the difference between El Niño and neutral events. Dotted black line: histogram of the difference between La Niña and neutral events. All the differences pass the Kolmogorov-Smirnov test at the 95% level

The effect of El Niño and La Niña events on the length of wet spells (LWS) is less marked than their effect on the number and the intensity of wet spells. Indeed, the distributions of the two samples are very close except for the one-day wet spells which are less frequent during El Niño events. However, when the El Niño events of 1982 and 1987 which are not associated with positive anomalies in seasonal rainfall amounts (Figure 4) are excluded from the sample, the shape of the distribution shows a negative departure for the one-day wet spells and positive departures for the two and four–eight-day-long wet spells. This suggests that El Niño tends to be associated with longer wet spells than during La Niña.

The above analyses described wet spells properties during El Niño and La Niña events. Associated atmospheric dynamics fields during wet spells are explored now in the next section.

3.3. Anomalies in rainfall and atmospheric dynamics fields during a wet spell: El Niño versus La Niña years

Analyses performed in this section are based on a selection of wet spells. First, for each wet spell occurring at a given station, we retrieve its starting date, ‘s0’. Then, we keep only those wet spells for which 20% of the stations at least have the same ‘s0’ (i.e. they experience a synchronous start of a wet spell). This selection aims at filtering out the smallest events (note that results do not change using a threshold set at 15 or 25% of stations). The wet spells selected account for about 25% of the total number of wet spells. This represents ∼19 wet spells per year, and this number is not significantly higher during El Niño events than during La Niña events (at 95% according to the Student's t-test). Once the wet spells and their associated ‘s0’ dates are selected, we extract the period spanning from s0 minus 2 days to s0 + 1 day (‘s–2’, ‘s–1’, ‘s0’, and ‘s1’). We then build two samples averaging the days ‘s–2’ and ‘s–1’, and ‘s 0’ and ‘s 1’ to document the rainfall and the atmospheric dynamics fields before, and during the wet spell, making a distinction between El Niño, La Niña and neutral events. The same analyses performed considering ‘s–1’, and ‘s 0’ only give very close results.

The mean evolution from s–2:s–1 to s0:s1 of rainfall and associated atmospheric dynamics (U, V, Z) at 500 hPa is displayed in Figure 9. Note that given the season we study, based on how we selected our wet spells, and due to the large number of raingauges available, light rainfall (1 mm/day) are likely to occur over the network at s–2:s–1 (Figure 9 (a)). Comparing patterns s–2:s–1 (Figure 9 (a) and (b)) with s0:s1 (Figure 9 (c) and (d)), there is a clear increase in rainfall amounts during wet spells associated with the arrival of the low pressure system, with amounts of 10 mm/day and more in Cape Town area and the seaward sides of the Hottentots Holland, Drakenstein and Cederberg mountains. This is due to the role of the mountain escarpments that uplift the air mass but it could also be due to the uneven distribution of stations over the domain with Cape Town area being overrepresented.

Figure 9.

Left panels: mean rainfall field before (s–2:s–1) and during (s0:s1) a wet spell. The circles size is proportional to the rainfall amount (in mm). Right panels: mean atmospheric dynamics (U, V and Z at 500 hPa) field before (s–2:s–1) and during (s0:s1) a wet spell (in standard deviation). This figure is available in colour online at

The differences in rainfall (in mm) between El Niño and neutral, and La Niña and neutral events, for the s–2:s–1, and s0:s1 periods are presented in Figure 10. The significance of the difference is tested according to the Student's t-test. It is clear that El Niño and La Niña events do not have perfectly symmetrical impacts. During El Niño events (Figure 10 (a) and (b)), a wet spell (s0:s1) is associated with widespread positive anomalies of rainfall, but these anomalies are mainly significant for the stations in Cape Town area and along the Hottentots Holland and Drakenstein escarpment. During the two days preceding the start of the wet spell (s–2:s–1), there is a clear opposition on each side of 33°S with stations to the north (south) experiencing negative (positive) anomalies of rainfall. La Niña events (Figure 10 (c) and (d)) feature different patterns. First, the patterns before and during wet spells are roughly the same with a clear opposition between stations on each side of 20°E: negative (positive) anomalies are recorded to the west (east). Moreover rainfall amounts are significantly modulated (low amounts) in many more stations than during El Niño events. Stations located north of 33°S are particularly affected. This suggests that either rain-bearing systems are shifted to the south during La Niña events or that they are of a different nature and bring less rain than during El Niño events.

Figure 10.

Differences (in mm) between El Niño and neutral events (top panels), and La Niña and neutral events (bottom panels) in rainfall field before (s–2:s–1) and during (s0:s1) a wet spell. Red (blue) circles denote negative (positive) differences. The circle size is proportional to the difference. Orange (blue) face colours denote differences significant at the 95% level according to the Student's t-test

The same analyses are carried on the atmospheric dynamics (U, V, Z, Figures 11 and 12). The significance of the differences between El Niño and La Niña events is tested according to the Student's t-test for Z and the Hotelling T-square statistics (Hotelling, 1931) for U and V. Patterns at 850 and 500 hPa are quite similar, hence, we display the 500 hPa ones only. During El Niño events and as compared to neutral events (Figure 11), the atmospheric dynamics accompanying wet spells is characterized by deeper lows embedded in the westerly flow. The cyclonic circulation centred at 0–40°S (Figure 9 bottom right panel, and Figure 11 (a)) is stronger with anomalies of divergence ahead of the trough (18°E/28°S) before the start of the wet spell (s–2:s–1). This induces a stronger than normal northwesterly flow offshore from s–2 to s–1. These atmospheric dynamics patterns are in agreement with rainfall patterns presented in Figure 10. In particular, before the start of wet spells (s–2:s–1), the anticyclonic circulation anomaly centred on the Namibia border is prone to inhibit rainfall which explains the negative rainfall anomalies recorded from Lamberts Bay to the Namaqualand. On the contrary, the Cape Town area and the seaward sides of the Hottentots Holland and Drakenstein mountains directly exposed to the northwesterly flow, tend to record slightly higher rainfall amounts. During the wet spell (s0:s1, Figure 11 (b)), all stations are under the influence of stronger than normal northwesterlie winds which are consistent with the widespread positive anomalies of rainfall. In addition (not shown), the 850 hPa temperatures over our study region and offshore tend to be warmer before the start of the wet spell then colder during the wet spell. These atmospheric dynamics patterns are typical of low-pressure systems such as cold fronts or cutoff lows which are known to bring rainfall over the western part of South Africa in winter (Preston-White and Tyson, 1997). This pattern is strengthened during El Niño events.

Figure 11.

Differences between El Niño and neutral events in wind (arrows) and geopotential height (contours, in hPa) fields at 500 hPa before (s–2:s–1) and during (s0:s1) a wet spell. Shadings (black arrows) denote differences significant at the 95% level according to the Student's t-test (Hotelling T-square statistics)

Figure 12.

Same as Figure 11 but for differences between La Niña and neutral events

The atmospheric dynamics pattern of anomalies during La Niña events (Figure 12) is different. Before the start of the wet spell (s–2:s–1, Figure 12 (a)), strong cyclonic anomalies are found off the south coast, a deeper low is present at 0–40°S, and a weaker high around 30°E–35°S (Figure 9 (c)) associated with stronger than normal westerlies wind above the western region. However, at 850 hPa (not shown), winds anomalies are rather southwesterly with anomalies of divergence over our study region. This could explain the significant negative rainfall anomalies recorded from Cape Town to the Namaqualand (Figure 10 (c)). During wet spell (s0:s1, Figure 12 (b)), in conjunction with high pressure anomaly to the west (∼10°E) and low-pressure anomaly to the east (∼40°E), strong southerly wind anomalies are observed over the region. The Garden Route which is windward, records positive anomalies of rainfall while the area from Cape Town to Namaqualand, which is downwind, records negative anomalies of rainfall (Figure 10). Cooler air is brought over the region as well (T, not shown). This atmospheric dynamics configuration is close to the southerly meridional flow described by Preston-White and Tyson (1997) that is known to bring rainfall to the Garden Route due to the orography.

3.4. ENSO—rainfall lead-lag relationships

Sections 3.1. and 3.2. have shown moderate positive (0.4–0.6) synchronous correlations between N3.4 and MJJ seasonal rainfall amount, number of Wet Spells (NWS) north of 32°S, and intensity of Wet Spells (IWS) at the raingauge scale. This raises the question of the level of skill rainfall forecasts based on ENSO pre-season values or ENSO forecasts could have. Following the approach proposed by Korecha and Barnston (2007) in their study of the predictability of the June–September rainfall in Ethiopia, Figure 13 displays correlation between the Niño3.4 index and seasonal amount (SA), NWS, LWS and IWS from 10 months before the rainy season to 9 month after (i.e. Niño3.4 in August year − 1 correlated with SA in MJJ year 0 and Niño3.4 in April year + 1 correlated with SA in MJJ year 0). Correlations are computed considering a rainfall regional index (average of the standardized 682 raingauges). Indeed, we expect stronger correlations than for any individual raingauge, due to the filtering effect of the spatial aggregation with respect to the random variability present in single raingauges.

Figure 13.

Lead-lag correlation (from August year − 1 to April year + 1) between the Nino3.4 index and the regional index of MJJ (year 0) seasonal amount (SA, plain line), number of wet spells (NWS, dashed line), length of wet spells (LWS, dashed-dotted line) and intensity of wet spells (IWS, dots). Circles denote correlations significant at 95%

The MJJ seasonal amount correlation with N3.4 in preseason (September to April, Figure 13) increases gradually with more than 25% of common variance from March (r> = 0.5) and correlations significant by January. During the season, the highest correlation is with Niño3.4 in May, then correlations decrease and are no longer significant. For the intra-seasonal components NWS, and LWS correlations are most of the time weak and not significant, even during the rainy season. This is not surprising given the few raingauges for which NWS and LWS are significantly correlated to N3.4 (Figure 6). IWS features a different pattern. Correlations are significant mainly during and after the rainy season suggesting that IWS could be a potential predictor of ENSO and not the reverse. This is coherent with recent findings by Jin and Kirtman (2009) and Terray (2011) who show that ENSO forced pattern in the Southern Hemisphere extra-tropics leads the peak phase of ENSO by one season. While the former authors attribute it to divergence/convergence anomalies sensitivity to local seasonality, the later suggests the setup of a subtropical forcing from the Atlantic and Indian basins on the equatorial Pacific. It involves the propagation of SST anomalies from the Atlantic and Indian subtropics into the tropics from boreal winter to spring then a dynamical atmospheric response over the three tropical basins which might trigger the onset of ENSO.

On the whole, these results suggest that the use of Niño3.4 preseason values would lead to a poor forecast of winter rainfall. The within season correlation values although higher than the preseason ones, remain moderate with the highest values found in May for SA and LWS and in June for IWS. Looking at Table I which presents Niño3.4 autocorrelation, it is obvious that the June–July state of Niño3.4 can be hardly inferred from its state the preceding months, while the May state can be accurately inferred from April. This weakening of the ENSO autocorrelation in spring is known as the ‘spring predictability barrier’. So we expect the use of either Niño3.4 preseason or forecasted values to be of moderate utility and skill to forecast the MJJ rainfall amount or intra-seasonal components.

Table I. Niño3.4 index autocorrelation (×100) over the period 1979–1999. Correlation values above 0.6 between season (May–July) and preseason (January–April) are in bold. Correlations below 0.2 are not displayed
Apr 191624469
May  1806687
June   19398

4. Discussion and conclusion

The aim of that study was to investigate the existence of an ENSO signal in the winter rainfall region of the western part of the Republic of South Africa. The results point out a significant association between ENSO and winter rainfall from the so-called 1976/1977 climate shift from May to July, peak of the annual cycle of rainfall there. This relationship has never been reported in previous studies partly because they usually considered the May–September or July–September season or longer periods (1900, or 1950–2000). Moreover whereas the statistical significance of the decadal variations of rainfall teleconnections with ENSO is questioned by Van Oldenborgh and Burgers (2005), the impact of the ENSO detected here features a strong decadal component and seems restricted to the recent decades. The correlation between MJJ rainfall amount 1st PC and the Niño3.4 index over 1979–1999 equals 0.49 (∼25% of common variance). In particular, 5 cold (warm) events out of 7 are synchronous with below (above) normal rainfall amounts. There seems to be no difference in rainfall anomalies according to the phase of ENSO (onset or decay).

The availability of daily rainfall data over a dense network of raingauges (682 stations extracted) allowed us to examine changes in wet spells properties during El Niño and La Niña events. Typically, positive (negative) anomalies in winter (MJJ) rainfall amount during El Niño (La Niña) events are due to wet spells which are longer (shorter) and bring more (less) rainfall. A majority of stations experiences a significant increase (decrease) of wet spells daily rainfall amounts ranging between 10 and 50 mm: the correlation between the Niño3.4 index and the frequency of wet days jumps from 0.27 to 0.56 when only those wet days recording more than 10 mm are considered. Following Reason et al. (2002), the hypothesis of rainy systems shifted to the north during El Niño events is proposed first to explain these results. Indeed, wet spells are more frequent over the northern stations and the probability of occurrence of a wet day around a wet station increases significantly for distances comprised between 480 and 720 km. The analyses carried on atmospheric dynamics (geopotential, wind and temperature) associated with the largest wet spells suggest that features of rainy systems affecting the western region seem involved as well. The 500 hPa geopotential and wind anomalies during El Niño events indicates a pattern usually observed during high and widespread rainfall events triggered by deep lows and associated cold fronts (Preston-White and Tyson, 1997) whereas during La Niña events, the flow has a stronger than normal southerly component which favours rainfall along the Garden Route and inhibits rainfall further north. Following Fauchereau et al. (2009) or Pohl et al. (2009) who have analyzed the different patterns of convection, the associated rainfall anomalies and the correlation with ENSO in summer, further work will focus on weather types themselves to confirm if their occurrence in winter is significantly modulated during El Niño and La Niña events.

The lack of signal in rainfall during the 1985 cold events, and the 1982 and 1987 warm events seems independent from the ENSO life cycle (onset or decay) or type. It could result from interferences with other climate modes. Moreover, the analysis of lead-lag correlations between the Niño34 index and various rainfall properties suggests that the regional seasonal amount or intra-seasonal components can be hardly forecasted from Niño3.4 preseason or forecasted values, the skill being too low.

Beyond these results, our study demonstrates the usefulness of analysing intra-seasonal rainfall characteristics to fully understand what causes changes in seasonal amounts and apprehend the scale interactions. It demonstrates, also, the necessity of treating El Niño and La Niña events separately at the intra-seasonal scale: in our case, computing differences between El Niño and La Niña events masks the southerly wind anomalies which lead to the hypothesis of a modulation of the weather types during El Niño and La Niña events. Lastly, following Ropelewski and Bell (2008) our study also calls for development and regular update of daily or hourly rainfall databases at the station-scale which are the only ones giving access to the rainy seasons intra-seasonal characteristics.


Dr N Philippon is grateful to the French National Scientific Research Centre (CNRS) who allowed a 10-month scientific visit to the Department of Oceanography of the University of Cape Town, and to that latter who hosted her. This study was part of the Program of International Scientific Cooperation PESOCA, co-funded by France and South Africa. Mathieu Rouault thanks NRF, WRC, ACCESS and Nansen-Tutu Center for funding. This is a contribution to the SEACHANGE NRF project and the SATREPS/JICA ACCESS Climate Program.