Linear and nonlinear statistical analysis of the impact of sub-tropical ridge intensity and position on south-east Australian rainfall


  • Kirien Whan,

    Corresponding author
    1. The Fenner School of Environment and Society, The Australian National University, Canberra, Australia
    2. The Climate Change Research Centre, The University of New South Wales, Sydney, Australia
    • Correspondence to: K. Whan, The Climate Change Research Centre, Level 4, Matthews Building, The University of New South Wales, Sydney 2052, Australia. E-mail:

    Search for more papers by this author
  • Bertrand Timbal,

    1. The Centre for Australian Weather and Climate Research, Bureau of Meteorology, Melbourne, Australia
    Search for more papers by this author
  • Janette Lindesay

    1. The Fenner School of Environment and Society, The Australian National University, Canberra, Australia
    Search for more papers by this author


The intensity and position of the sub-tropical ridge (STR) have strong relationships with rainfall variability in southern Australia. The combined effect of intensity and position in March-April-May (MAM) and June-July-August (JJA) is the focus of this research. Linear statistics were used first: area-averaged and Australia-wide spatial correlations of STR intensity and position with precipitation in south-west eastern Australia reveal that STR intensity has a much stronger and more widespread relationship with precipitation in both seasons. Over time, these relationships vary in magnitude and spatial extent with the sign of the correlation changing between two 50-year epochs. These nonlinearities were investigated further using classification trees. Area-averaged precipitation data (terciles) for south-west eastern Australia was classified on the basis of STR intensity and position. In both seasons the classification trees identify STR intensity as the primary partition defining the dry group, supporting the linear analysis. In the transition season of MAM, the time of year when the mean position of the STR is more southerly, STR position is important in distinguishing between a ‘winter-like’ and a ‘summer-like’ wet groups, providing STR intensity is low. Vector wind analyses were computed to explain the composite seasonal precipitation anomaly results in terms of different circulation patterns associated with these two wet groups. The frequency of wet and dry cases in each group was examined with changes evident over the recent years. The research confirms that STR intensity is more important than STR position in explaining inter-annual rainfall variability across southern Australia but also demonstrates the additional role of STR position in MAM. These results explain the low correlation between rainfall and STR position and why this relationship has evolved during the 20th century as the mean location of the STR has shifted south in MAM.

1. Introduction

Since 1997 south-eastern Australia has experienced record low rainfall, exacerbated by record high temperatures (Murphy and Timbal, 2008). Several large-scale modes of climate variability have relationships with south-east Australian climate and may therefore have contributed to the recent rainfall deficit, including tropical modes of variability; such as El Niño-Southern Oscillation (ENSO) (Nicholls et al., 1997, Trenberth, 1997, Suppiah and Hennessy, 1998) and the Indian Ocean (Simmonds and Rocha, 1991, Saji et al., 1999, Ummenhofer et al., 2011); high-latitude modes such as the SAM (Marshall, 2003, Meneghini et al., 2007) and regional mean sea level pressure indices such as the sub-tropical ridge (STR) (Drosdowsky, 2005, Larsen and Nicholls, 2009, Williams and Stone, 2009, Timbal and Drosdowsky, 2013). The spatial and temporal signature of the recent rainfall deficit makes the STR particularly relevant for two reasons. Firstly, the majority of the deficit is evident in MAM (March-April-May) when the tropical modes have little influence (Timbal, 2009) and secondly, the spatial extent of the deficit matches well with the area under the influence of the STR (Timbal and Drosdowsky, 2013).

The STR is the downward branch of the Hadley Cell; a band of high pressure that circles the globe at approximately 30° latitude in both hemispheres. The subsiding air associated with this band of high pressure is responsible for arid regions on several continents, including Australia (Peixoto and Oort, 1992). Increasing global temperature is associated with changes to the general circulation of the atmosphere including changes to the Hadley Cell. Generally the Hadley Cell has been broadening, leading to a poleward expansion of the sub-tropical dry zone (Seidel et al., 2007, Zhou et al., 2011). However, there is little scientific consensus about changes to the Hadley Cell, as disagreement exists between reanalysis products and global climate models. A study using precipitation observations supports the results from reanalysis products and points to a poleward shift in the sub-tropical dry zone (Zhou et al., 2011). In addition, regional variability exists in these trends; in particular, the sub-tropical dry zones associated with the downward branch of the Hadley Cell have expanded in regions under the influence of the Asian monsoon. Hemispheric differences are also evident as larger significant trends are found in the Northern Hemisphere, while the Southern Hemisphere displays insignificant trends towards Hadley Cell expansion (Zhou et al., 2011). Changes in the Hadley Cell are associated with changes to the STR, it is therefore imperative that we understand how changes in the characteristics of the STR may affect the hydrometeorology of south-east Australia.

The STR is calculated over eastern Australia (Drosdowsky, 2005, Timbal and Drosdowsky, 2013) and is closely linked to precipitation variability in south-east Australia (Timbal and Drosdowsky, 2013). The seasonal cycle of the STR (Figure 1) shows that ridge position is furthest south in late summer (39.6°S in February) and furthest north in late winter (29.1°S in August). Intensity is highest from April to September, reaching a peak of 1022.5 hPa in June. Summer experiences the weakest STR intensity with a minimum of 1013.3 hPa reached in December. The STR moves north during MAM, with the largest northward shift from March to April as it moves an average of almost 3° over the month. The STR is north of south-east Australia in June-July-August (JJA) and September-October-November (SON) and south of south-east Australia in December-January-February (DJF) and MAM. Thus, the STR is more intense and at lower latitudes during JJA compared with that in MAM. Some differences are evident between the seasonal cycle presented here and that used in other research because different periods are used (Drosdowsky, 2005, Cai et al., 2011).

Figure 1.

The monthly averages showing the annual cycle of mean STR position and intensity. Black – months averaged from 1900 to 2009, blue – months averaged from 1950 to 1979 and red – months averaged from 1980 to 2009.

The seasonal north–south movement of the STR plays an important role in determining the climate of southern Australia through changes to the mid-latitude storm track (Murphy and Timbal, 2008). When the STR is at higher latitudes in DJF, westerly winds, with embedded frontal activity, are pushed southwards missing continental Australia, while tropical influences may bring precipitation to the region. When the STR is at lower latitudes in JJA, cold fronts are able to pass over south-eastern Australia bringing rain to the region. As south-eastern Australia receives the majority of its precipitation in JJA, the importance of the meridional movement of the STR is obvious, and research has shown that up to 60% of inter-annual variability in JJA rainfall is related to movement of the STR (Williams and Stone, 2009). The intensity of the STR is also related to rainfall, as increasing pressure is associated with decreasing rainfall in the sub-tropics (Ansell et al., 2000). This study will expand our understanding of how inter-annual variability in STR intensity and position combine to affect rainfall.

Historically, research has focused on how the variability in STR position influences climate, particularly in the east of Australia (Pittock, 1973, Drosdowsky, 2005, Williams and Stone, 2009). Research into the recent rainfall deficit shows that a majority (62%) of the deficit can be linked to the intensity of the STR (Timbal and Drosdowsky, 2013); using simple linear statistics, Timbal and Drosdowsky (2013) were unable to find an effect of STR position on rainfall in addition to the effect of STR intensity. They noted that overall the relationships between the two aspects of the STR are fairly stable overtime (as are the relationships between the two facets of the STR and rainfall, in particular for the STR intensity and rainfall relationship). However, other researchers have found changes in the relationship over time (Drosdowsky, 2005, Cai et al., 2011). Cai et al. (2011) showed that the STR intensity and position relationship has weakened during the last 30 years, particularly in MAM. Increasing understanding on whether the MAM relationship between STR intensity and position is changing, and if so increasing understanding about the nature of that change, is part of the motivation for this paper. An additional motivation is to understand why STR position does not have an effect on rainfall in addition to that of intensity, despite the fact that STR position on its own can be linked to a small amount of the rainfall deficit. The lack of an additional effect attributable to the STR position is even more puzzling when considering that the delay in the northward shift of the STR in the MAM, evident since 1975, is the largest shift, and it is the season when most of the rainfall deficit has been observed record (Drosdowsky, 2005, Timbal and Drosdowsky, 2013).

Globally, as many climatological parameters display a significant climate shift from the mid-1970s (Trenberth, 1990) it is interesting to also see this shift in the STR record. Figure 1 shows the shift in STR intensity and position from the early (1950–1979) to the later (1980–2009) part of the record, with the largest changes found in the transition seasons (MAM, SON). In the latter period the STR is more intense and further south in April, May, June and November, while STR intensity is greater in July and August, compared to the earlier period.

Many authors have raised concerns about the use of linear statistics to describe climate processes that are intrinsically nonlinear (Zorita et al., 1995, Campbell et al., 2000, Firth et al., 2005). Linear statistics are a useful tool to help us understand the climate system but are one tool of many that researchers should employ and have a limited ability to deal with the complexity of the climate system. Nonlinear techniques should be used in conjunction with traditional methods to help us fully understand relationships and interactions between complex variables, such as precipitation, that have multiple nonlinear associations with highly correlated predictor variables. Classification and regression trees (CART) is one nonlinear strategy that appears to be appropriate for the problem described. CART is able to provide insight into the physical mechanisms underlying statistical relationships with easily interpretable results and is shown to have the greatest advantage over linear techniques when applied to data that is nonlinear (Rodionov et al., 2001).

The importance of the STR for southern Australian rainfall is clear. Changes in STR intensity have a verified effect on precipitation (Ansell et al., 2000, Murphy and Timbal, 2008, Larsen and Nicholls, 2009, Timbal and Drosdowsky, 2013), yet the role of the position is less well understood. This study will explore what effect inter-annual variability in STR position has, on south-east Australian rainfall (in addition to that of intensity), using both linear and nonlinear statistical techniques. The next section will describe the datasets and methods used, including an introduction to the CART method. The following sections will present the results from both the linear and CART analyses. Then, composites of precipitation and vector wind are used to describe the atmospheric circulation associated with particular phases of the STR, based on results of the CART analysis. Finally, results are discussed further and conclusions are drawn on how STR intensity and position are associated with precipitation in Australia.

2. Data and methods

2.1. Datasets

This study relies on Australia-wide precipitation data sourced from the Australian Water Availability (AWAP) dataset (Jones et al., 2009) from 1900 to 2009. The AWAP dataset interpolates station data and is available on a 0.05° grid for the Australian continent. A precipitation time series for south-west eastern Australia (SEA) was calculated by seasonally averaging values within the hatched region in Figure 2. This region is the area where the STR influence is the greatest as STR intensity explains over 20% of the rainfall variability. SEA approximately covers the area to the south-west of a line running from Melbourne to the south of the Flinders ranges and following the end of the Great Dividing Range (GDR) over western Victoria (see Timbal et al., 2010 for detail about the construction of the SEA time series). The 110-year SEA precipitation time series was divided into three groups, with 37 cases in the dry and average groups and 36 cases in the wet group.

Figure 2.

Map of south-east Australia. The hatched area defines the south-west of eastern Australia (SEA) which is area-averaged and used in this study.

The STR intensity and position indices are calculated as per the station-based ‘L-index’ in Drosdowsky (2005). The indices are based on monthly mean-sea level pressure station data between 10 and 44°S along the eastern Australian coast. The station data is interpolated to 1° grid and averaged within a 5° longitude band around 150°E. The location of the maximum in pressure within that band gives the position of the STR (with 1° accuracy), as well as the intensity of that maximum. This index is simple to calculate, robust and consistent over time (Drosdowsky, 2005). The use of station data means that the index can be calculated as far back as 1890; however, in this study we used data from 1900 to 2009 to match the available gridded precipitation data. ENSO years are obtained from the Bureau of Meteorology list of 25 ENSO years (Table 1) that have occurred since 1900 (Bureau of Meteorology, 2010).

Table 1. ENSO years (from Bureau of Meteorology, 2010)
ENSO years
1902–1903, 1905–1906, 1911–1912, 1913–1914, 1914–1915, 1919–1920, 1925–1926, 1940–1941, 1941–1942, 1946–1947, 1951–1952, 1957–1958, 1963–1964, 1965–1966, 1972–1973, 1977–1978, 1982–1983, 1987–1988, 1991–1992, 1993–1994, 1994–1995, 1997–1998, 2002–2003, 2006–2007, 2009–2010

Monthly vector wind anomalies are obtained for the Australian region from the NCEP/NCAR (National Centers for Environmental Prediction/National Center for Atmospheric Research) Reanalysis (Kalnay et al., 1996). Seasonally averaged, the data is available from 1948 to present. The 850 hPa level was chosen to represent wind variability in the lower troposphere and may be able to provide some insight into precipitation; bearing in mind that rainfall is related to moisture convergence and that while 850 hPa is an important level for lower tropospheric moisture it does not capture the entire tropospheric moisture content nor does it include the effect of transient synoptic systems.

2.2. Methods

The first analysis is a linear examination of SEA precipitation and the two facets of the STR. Correlations (significant at the 95 percent level with a two-tailed Student's t-test) were conducted between the SEA area-averaged time series and indices of STR intensity and position, followed by spatial correlation analysis between continental gridded Australian rainfall (0.05°) with both facets of the STR. Statistical significance is defined throughout this research at the 95th percentile, which based on 110-years of data means correlations that exceed 0.2 are significant (and are thus coloured in Figures 4 and 5). A decrease in the degrees of freedom stemming from persistence in Australian rainfall (Simmonds and Hope, 1997) has not been included in the significance testing.

To explore the non-stationarity of the relationship between the STR and rainfall, and their spatial features, spatial correlation maps were produced for two 50-year epochs (Epoch 1: 1900–1949 and Epoch 2: 1960–2009) to contrast how these relationships have evolved during the 20th century but using long enough periods to ensure results are statistically significant. The same significance level is also used in the epoch analysis, despite the smaller sample size (which would mean that correlations significant at the 95th percentile exceed 0.28, rather than the 0.20 shown here). A common significance level is used here for both phases of the analysis because the purpose is to show changes in the relationships over time; hence to emphasize these temporal changes, the significance level is kept constant throughout the whole research. As per the previous discussion, some auto-correlation may exist in the JJA series so there may be some reduction in their significance. However, the focus is on MAM and the relative changes in the epochs. Previous research has shown the stability of STR-rainfall relationships in SEA over the instrumental record (Timbal and Drosdowsky, 2013), which gives confidence in the correlations shown in this region, although the same confidence cannot be given for those correlations outside the region (e.g. Queensland).

Next is the CART analysis, which uses area-averaged SEA rainfall time series as a categorical response variable (the three groups are dry, average and wet), with STR position and intensity as predictor variables. The cases that comprise each terminal node are used in a composite anomaly analysis of the AWAP precipitation data with anomalies calculated against the full record. The 20 most recent cases in each terminal node that occur after 1948 are also used in a composite anomaly of vector wind using the NCEP/NCAR Reanalysis product.

CART is a binary recursive partitioning technique first developed in the 1980s (Breiman et al., 1984). CART has been employed in many fields including ecology and genetics (De'ath and Fabricius, 2000, Thuiller et al., 2003, Esther et al., 2010), with some applications in climatology (Burrows et al., 1995, Rodionov et al., 2001, Firth et al., 2005, Rodionov et al., 2005). Decision trees are constructed that seek to describe the variability of one response variable by several predictor variables (De'ath and Fabricius, 2000). They are particularly useful when there are multiple correlated predictor variables, such as is the case with STR intensity and position in both MAM and JJA (r = 0.49 and 0.70, respectively). Trees may be constructed using either continuous (regression trees) or categorical (classification trees) response data; in this research classification trees are used (illustrative example in Figure 3). In the example, a response variable (n = 30) containing three classes (A, B and C) is classified on the basis of two predictor variables (1 and 2). This can be compared to the current research in which the response variable used is SEA precipitation (n = 110), which is divided into terciles (dry, average and wet groups), and classified on the basis of two predictor variables (STR intensity and position). The first parent node (Node 1), in the example, contains 30 cases of the three classes of response variable each with ten members. The daughter node to the left is named by doubling the parent node (i.e. Node 2), while the right-hand daughter node is named by doubling the parent node and adding one (i.e. Node 3). Splits are selected on the basis of predictor variables and continue until a terminal node (a node that has no further splits). This is can be seen in the example where the original parent node is ‘Node 1’ which is split into two daughter nodes (Terminal Nodes 2 and 3) on the basis of ‘Predictor variable 1’. When the splitting criterion is met (i.e. ‘Predictor variable 1’ is > x) the cases of the response variable are put into the left node. If the criterion is not met (i.e. ‘Predictor variable 1’ is ≤ x) the cases are put into the right node. Increasingly homogenous groups are created as parent nodes are split into daughter nodes on the basis of specific values of predictor variables. Splits are selected to maximize the increase in node purity, as measured by the Gini index. The Gini index measures node impurity and is calculated from of the number of cases belonging to each class in the node. The index is calculated by subtracting the sum of the squared probability of each class' occurrence in the node from one (Equation (1)).

Figure 3.

Classification tree example. Predictand data = three groups each with ten members (n = 30), which is classified on the basis of two predictor variables (predictor variables 1 and 2).

Figure 4.

Seasonal correlation coefficients between STR intensity (top row) and position (bottom row) and Australian rainfall in MAM, JJA, SON and DJF over the period 1900 to 2009. Shaded areas show significant at the 95% confidence level. Refer to Figure 2 for the SEA region.

Figure 5.

Correlation coefficients between STR intensity (top row) and position (bottom row) in MAM (left two columns) and JJA (right two columns) over two 50-year epochs: 1900–1949 (Epoch 1) and 1960–2009 (Epoch 2). Shaded areas show significant at the 95% confidence level. Refer to Figure 2 for the SEA region.

The Gini index, where j is the  class and pj is the probability of class j:

display math(1)

This means that when a node contains only one class the Gini index is zero, while when there are equal numbers of all classes the index is one. In the example, Node 1 has equal numbers of all classes so the Gini index is at its maximum value of one as node impurity is at its maximum. In Terminal Node 7 the Gini index is 0.26 showing that it is lower when node purity is higher.

Complex trees can be grown that maximize node purity by allocating every case to a separate terminal node, but such an overfit tree is unlikely to be replicable with new datasets, limits interpretability and has little practical significance. The result then is a trade-off between node purity and tree size or interpretability. An overly large tree is created and then pruned back based on an arbitrary decision, although in this research tree size is selected to minimize the cross-validated error and maintain physical consistency. In the cross-validation process, the dataset is randomly divided into subsets. To each subset of data a training/test sample approach is applied to estimate misclassification rates in sub-trees. From this, a complexity parameter is chosen that minimizes misclassification error and is used to select the optimal tree size. This enables tree construction that is both statistically robust and physically consistent. It is important to recognize that when looking at SEA rainfall using only STR intensity and position as predictor variables the highest values of node purity are not expected as this would fail to recognize the other modes of climate variability that are at play and are not being captured using only these two predictors. High node purity is easily achievable when tree size is large, but in this case the interpretability of the tree is often compromised. When tree interpretability is high, node purity is often conceded as the tree is pruned back to a smaller size more manageable size. Consequently, terminal nodes in classification trees are rarely completely pure (containing only one class of data) and are more regularly comprised of multiple data classes, as we would expect in climatic research using only two predictor variables to describe a complex variable such as precipitation. CART uses a ‘majority rule’ to allocate each terminal node to a class, so that the terminal node is allocated to the class to which a majority of its member belong. For example, in Terminal Node 2 there are seven cases in class A, four cases in class B and one case in class C. This means that this terminal node will be allocated as ‘class A’ and would be 58% pure as 7/12 cases belong to the class employed by the majority rule. CART is robust to outliers and extreme values among predictor variables as the technique tries to put divergent cases in separate nodes (Rodionov et al., 2001).

In the current study decision trees were constructed using the statistical package R (R Core Team, 2012), with the library ‘rpart’ (Therneau and Atkinson, 2010).

3. Results

3.1. Results from linear statistics

Over the full record, negative correlations between SEA precipitation and STR intensity and position (Table 2) show that as the intensity of the STR increases and as the STR moves further south, precipitation in SEA decreases.

Table 2. Correlation between precipitation in SEA and STR intensity/STR position

The correlation with STR intensity is stronger and more significant than that with STR position, particularly in MAM. This confirms previous linear research that found intensity is the most important aspect of the STR to explain SEA rainfall variability (Timbal and Drosdowsky, 2013), and is in general agreement with the known relationship between pressure and rainfall in the sub-tropics (Ansell et al., 2000).

The spatial correlation analysis of STR intensity and position with gridded precipitation is shown in Figure 4. The seasonal cycle of both facets of the STR and SEA rainfall are closely linked. Correlations strengthen during MAM and reach a maximum in JJA, before decaying throughout SON to a minimum in DJF. In MAM, the maximum correlation between rainfall and intensity (position) is −0.65 (−0.60). As the relationship develops into the JJA months, the maximum correlation between rainfall and intensity (position) is −0.74 (−0.68), again showing the dominance of intensity over position. The spatial extent of significant correlations for both intensity and position is the greatest in JJA, and larger for intensity than position. In MAM, all significant correlations with intensity are negative, so increases in STR intensity is associated with decreased precipitation (or no impact when the correlations are not significant). STR position, however, has a more spatially variable relationship with precipitation; negative correlations are evident in SEA but positive correlations are found further north. This shows that the relationship with position depends on the mean location of the STR, which in MAM varies between 37 and 32°S, in between the areas with significant positive and negative correlations. When the mean position of the STR is to the north (south) of the area considered, a shift south in STR position will be associated with decreased (increased) rainfall. The mean location of the STR in relation to SEA changes with the annual cycle (Figure 1), and also on a multi-decadal timescale (see later this section), as such, the complexities of the STR position–rainfall relationship are difficult to grasp with linear statistics. This is particularly relevant in MAM when the mean position of the STR is located above SEA and is changing rapidly from a southerly DJF position to a northerly position in JJA. The high variability found in the STR position record adds an additional layer of complexity to the issue and results in few significant trends (Timbal and Drosdowsky, 2013). However, this mechanism shows that subtle variations in the mean position of the STR are important for precipitation despite a lack of statistical significance found with linear statistics.

Temporal changes are found in the relationship between STR position and Australian rainfall over the instrumental record (Drosdowsky, 2005, Timbal and Drosdowsky, 2013). Timbal and Drosdowsky (2013) found variation in correlation coefficients (calculated over 30-year periods between 1929 and 2009) for relationships between SEA rainfall and both STR intensity and position. They found less multi-decadal variability in the correlation between SEA rainfall and STR intensity, with the largest changes evident in May. The relationship between STR position and rainfall, however, shows more variability between the two 30-year epochs (Timbal and Drosdowsky, 2013). Timbal and Drosdowsky (2013) found that in July correlations increase from −0.2 in the 1940s to −0.6 to −0.8 in the most recent period. In SEA, March and April had negative correlations during the early period, while the end of the record shows positive correlations in these months. These results outline the complexity of the relationship between the STR position and precipitation, particularly in MAM, the vital transition season between DJF and JJA patterns. Variability of the STR position is larger in JJA compared with that in MAM (standard deviation = 2.3 compared to 1.8). Yet the mean position of the STR in MAM is key, as the STR can be located either across or north or south of SEA in this season, with the nature of the relationship depending on the position of the STR. In the current research the epoch analysis (Figure 5) was conducted here to extend the understanding of STR non-stationarity. Looking firstly at the relationship between rainfall and STR intensity, the negative correlation is stronger and more extensive in Epoch 2, and when correlations are significant in both epochs, magnitudes are also greater in the later epoch. Specifically, in Epoch 2 significant correlations between STR intensity and rainfall are evident across all of New South Wales (NSW), the majority of Queensland and into the Northern Territory; compared with those in Epoch 1, where the significant correlations are confined to SEA and Tasmania. The multi-decadal differences associated with STR intensity in MAM include a reversal in the sign of the relationship as positive correlations present during Epoch 1 in Queensland are replaced by negative correlations in Epoch 2.

While the 20th century has seen an increase in the spatial extent and magnitude of the correlation between STR intensity and rainfall in MAM, the opposite is true in JJA. This season has experienced a reduction in the spatial extent of the correlation between STR intensity and precipitation from Epoch 1 to Epoch 2. However, the spatial reduction is not associated with a reduction in magnitude of the correlation over southern Australia where the relationship remains significant (in contrast to the increases in correlation over SEA in MAM). Epoch 1 displays a significant negative relationship over the majority of Australia (excluding Queensland), whereas in Epoch 2 the significant correlations are confined to southern Australia. This may be related to a small long-term southward trend in the position of the STR over the 20th century (Timbal and Drosdowsky, 2013). Although Timbal and Drosdowsky (2013) found the strength of the STR intensity correlation in SEA to be relatively steady over the 20th century, this analysis shows that there are spatial differences in the extent of the relationship (away from SEA) in both MAM and JJA.

Secondly, the relationship between STR position and precipitation has also changed between the two epochs. SEA exhibits a reversal in the sign of the correlation between position and rainfall over the 20th century in some months (Timbal and Drosdowsky, 2013), which is evident in the spatial correlation maps. In MAM during Epoch 1, there is a region of positive correlation in north Queensland with negative correlation covering most of SEA. The region of positive correlation shifted south and stretches east from South Australia down the east coast of NSW, while the areas with negative correlation shrink and are confined to a small fringe along the coast from South Australia to Victoria and western Tasmania in Epoch 2. During JJA, the spatial extent of the correlation decreases over the 20th century, while the magnitude and sign of the correlation over SEA are broadly unchanged. The key features are a negative relationship dominating the SEA region and extending along the coast to Western Australia and a positive relationship on the east coast.

This section of the analysis highlights the importance of utilizing long records in research, as even here with 50-year epochs the non-stationarity of the series is evident. Different conclusions can be drawn when a smaller or longer subset of the record is selected for the analysis, although this can highlight how relationships change over time. Cai et al. (2011) concluded that MAM rainfall cannot be linked to the STR as only insignificant correlations (−0.35 and 0.22 for intensity and position respectively) can be found in this season. This difference stems from using a smaller subset of the dataset, although when longer time periods are examined the relationship is significant (Table 2). The positive correlation (Cai et al. 2011) found between STR position and SEA rainfall from 1979 to 2008 reflects the shifting sign of the relationship found by Timbal and Drosdowsky (2013) and shown spatially in Figure 5.

It is worth noting that for SEA overall in JJA the evolution of the relationship between STR intensity and position with rainfall between Epoch 1 and Epoch 2 is broadly similar, with the biggest changes evident further north over inland Australia with little change across SEA. This suggests that while the STR has shifted south (Timbal and Drosdowsky, 2013), it has not shifted south enough to change the relationship with rainfall over the southern parts of the continent because of the northerly mean position of the STR in this season. Most changes are evident further north where a shift south of the STR has had sizeable impact. More interesting are the changes in the relationships, outlined previously, between rainfall and both facets of the STR in MAM; most notably, a strengthening of the relationship with intensity over most of eastern Australia including SEA and a contraction of the negative influence of the STR position to a very small part of SEA in conjunction with the increase in the area with a positive relationship. These results illustrate that the relationship between the position of the STR and rainfall is highly complex and changes over time depending on the mean location of the STR.

The association between the STR and SEA climatic variability is therefore evidently nonlinear with the impact dependent on the background climatology (i.e. the difference in the mean between MAM and JJA). The current linear analysis confirms and extends recent research and gives a good understanding of how the STR varies throughout the year and how it has changed over the instrumental record. Linear analysis has confirmed the dominance of STR intensity (compared with position) in controlling SEA precipitation but does not provide a clear picture about the interaction between the two aspects of the STR. Therefore, while linear statistics are useful in basic description and analysis of such nonlinear processes, it is essential that more complex statistics are employed to fully understand STR interactions. We will now use CART to explore the nonlinear interactions between intensity, position and SEA rainfall.

3.2. Results from CART analysis

3.2.1. MAM

In MAM, the CART procedure identifies four terminal nodes, with both predictor variables required to adequately classify the data (Figure 6). The first split uses STR intensity to separate 40/110 cases, confirming the predominance of STR intensity from the linear analysis. This split defines the dry Terminal Node 2, when high STR intensity ≥1021 hPa. Node purity is 60% as high pressure is associated with low precipitation in nearly 2/3 of the cases, a high percentage when considering the sample size (24 out of 37 dry years are identified solely on STR intensity being above 1021 hPa). Composite rainfall anomaly analysis of Terminal Node 2 (Figure 7) shows the spatial impact on precipitation as this node is characterized by negative precipitation anomalies over much of SEA, with the largest negative anomalies are located in the areas of largest mean rainfall (western Tasmania and on the Southern Alps). High pressure is associated with a reduction in the mean westerly flow (Timbal and Drosdowsky, 2013) and the effects of this are clearly seen in the spatial pattern emerging from the composite anomaly map.

Figure 6.

Decision tree from the CART analysis in MAM after pruning to four terminal nodes.

Figure 7.

Seasonal precipitation composite anomaly maps (mm/year) for a large part of the Australian continent encompassing SEA in MAM. Composite are based on seasonal means from 1900 to 2009.

The first split supports the proposition that STR intensity is more important than position in determining drought in SEA (Timbal and Drosdowsky, 2013); however, the position of the STR is necessary to understand the full range of precipitation variability as the second split (for cases below 1021 hPa) is made on position. When intensity is below 1021 hPa and the position of the STR is further south (≥36.17°) the CART technique classifies Terminal Node 7 as a wet node with very high purity; ten cases during 110 years are classified in this terminal node with 80% belonging to the wet class and no dry cases classified into this group. The composite analysis (Figure 7) of Terminal Node 7 displays positive rainfall anomalies over the whole of SEA including the northern inland regions with very high anomalies on the eastern coast. However, this very wet node displays a strip of negative rainfall anomaly over the south-west coast of eastern Australia and Tasmania with the largest negative anomalies on the western Tasmanian coast. This pattern of negative and positive rainfall anomalies is consistent with tropical penetration of moisture from northern Australia and is reminiscent of a ‘summer-like’ rainfall pattern, as opposed to the ‘winter-like’ rainfall anomalies where one would expect to see largest rainfall anomalies along the southern edge of the continent.

When the position of the STR is in a more northerly position (<36.17°) the third split is based on intensity. When the STR intensity is low (<1019 hPa) and the STR is in a northerly position a wet MAM occurs (Terminal Node 13). The composite seasonal precipitation anomaly for this terminal node (Figure 7) shows that positive precipitation anomalies confined to SEA, with the largest positive anomalies along the Southern Alps and in western Tasmania while negative anomalies are present in this node along the eastern coast of NSW. Spatially rainfall is consistent with an increase westerly flow and is therefore more ‘winter-like’. Although Terminal Nodes 7 and 13 are both classified as wet nodes, spatially the precipitation between these two nodes are very different. In Terminal Node 7 wet conditions occur when the STR is in a southerly position (≥36.17°) even when STR intensity is above average (up to 1021 hPa while the MAM mean is below 1020 hPa), while in Terminal Node 13 wet ‘winter-like’ conditions occur only when STR intensity is below average and the STR is in a more northerly position (<36.17°). These two terminal nodes show two of the possible positions of the STR in the transition season.

Finally, when the STR is at low latitudes and STR intensity is at median levels (between 1019–1021 hPa) an average terminal node is defined. Terminal Node 12 is a highly heterogeneous node with 11 dry cases, 17 average cases and 8 wet cases. The low node purity clearly indicates that this final node is poorly determined by both STR intensity and position, and hence in these cases another variable is likely to be the primary driver of the rainfall variability. This is logical because it corresponds to years with near average STR intensity and position. As expected, the rainfall composite map for Terminal Node 12 does not show spatially consistent rainfall anomalies across the SEA region although further south, in Tasmania, positive anomalies are similar to the rainfall regime associated with Terminal Node 13. The low latitude of the STR in this node is only evident in these Tasmanian anomalies while across the continent rainfall anomalies are small and patchy.

The years identified in each terminal node can give additional information about variability in rainfall regimes. Figure 8 shows the time series for SEA rainfall in MAM (vertical bars), the terciles that the data was split into for the CART analysis (horizontal lines) and the terminal nodes each year was classified into (colours of vertical bars). It is clear that the recent MAM-dominated rainfall deficit is associated with STR intensity. Eleven of fourteen years in the last two decades classified in Terminal Node 2 are dry years. This supports the research that found increased STR intensity to be the primary driver of the recent rainfall deficit, particularly in MAM (Timbal and Drosdowsky, 2013) as high pressure is associated with lower rainfall owing to increased stability of the air column and a weakening of the mean westerly airstream.

Figure 8.

SEA MAM precipitation time series. Colour of the vertical bar indicates the terminal node the year has been classified into. The red and blue horizontal lines indicate the cut-off points for the original classification (dry, average and wet), giving an indication of misclassification when the horizontal and vertical classifications do not match.

The ‘summer-like’ rainfall pattern associated with Terminal Node 7 has been a rare event over the instrumental record, with only ten cases recorded. The low number prevents an in-depth analysis but it can be noted that while only four cases were recorded by the first half of the record (1900–1955), six were recorded in the second half, with cases observed in every decade and hence are a continuing occurrence. On the contrary, while Terminal Node 13 has been more common over the entire record (25 cases recorded), in the last 30 years has only seen this ‘winter-like’ pattern occur once in 1995, while the ‘summer-like’ wet pattern was observed three times (1989, 1990 and 2007).

Overall, increasing STR intensity over the latter part of the 20th century has seen an increase in the occurrence of the dry Terminal Node 2, a near disappearance of the ‘winter-like’ wet Terminal Node 13 in the last 30 years and a small increase of the rare occurrences of ‘summer-like’ wet MAM.

3.2.2 JJA

In JJA, the first split in the CART model is on intensity (Figure 9), separating a dry node (Terminal Node 2) with 49/100 years (with 65% purity) when intensity is high (≥1022 hPa); this is very similar to MAM. The higher threshold value in JJA compared with that in MAM is consistent with the higher climatological value as shown in the annual cycle (Figure 1). The increased node purity in JJA's Terminal Node 2 compared with MAM is consistent with the higher correlation between STR intensity and rainfall in JJA compared with that in MAM. The composite rainfall anomaly analysis (Figure 10) for this high intensity group (Terminal Node 2) displays a broad region of negative precipitation anomalies over SEA, with the largest anomalies on the mountain regions and western Tasmania, combined with a small region along the eastern seaboard that displays a positive precipitation anomaly. Again, for SEA, this is very similar to MAM. The main difference between the two seasons is that in JJA the second split is also based on intensity (<1020 hPa), with position counting only in a third split between Terminal Nodes 12 and 13 (Figure 9). Terminal Node 7 is classified as wet and is 100% pure but is a rare occurrence (6/110 years). The composite anomaly pattern of this low intensity wet terminal node has a spatial pattern that is reminiscent of the ‘winter-like’ Terminal Node 13 in MAM with the largest positive anomalies along the Alps and in the west coast of Tasmania and the mainland with negative anomaly found on the eastern seaboard, i.e. east of the GDR. The rainfall patterns of Terminal Nodes 2 and 7, which are mirror images suggest a linear relationship between intensity and precipitation as high or low pressure (either ≥1022 or <1020 hPa respectively) is associated with high and low precipitation extremes.

Figure 9.

Decision tree from the CART analysis in JJA after pruning to four terminal nodes.

Figure 10.

Seasonal precipitation composite anomaly maps (mm/year) for a large part of the Australian continent encompassing SEA in JJA. Composite are based on seasonal means from 1900 to 2009.

The rainfall pattern emerging from Terminal node 12, highly heterogeneous modes with many average years as was the case in MAM and more wet years than dry years and a rainfall pattern tending towards wetter across SEA but half way between wet Terminal Node 7 and dry Terminal Node 2.

An interesting result of the JJA CART model is that when STR intensity is fairly neutral (between 1020 and 1022 hPa), the position of the STR is identified as being important. When the position of the STR is at further south (latitudes >30.5°) the result is Terminal Node 13, a wet node with high purity (77%). Although, the CART model was constructed using anomalies across SEA and the rainfall anomalies across SEA are very similar between Terminal Nodes 12 and 13, the most visible differences on the composite anomaly maps are located further north. Negative rainfall anomalies are evident across eastern Australia when the STR is further north (Terminal Node 12), but positive anomalies extend across all eastern Australia when the STR is located further south (Terminal Node 13). This difference between the two nodes illustrates that the two facets of the STR (intensity and position) are not independent; a ridge of moderate intensity will prevent rainfall bearing westerly front to penetrate inland is located north, but this effect disappears when the STR is displaced further south. Although it is also very likely that in those years with neutral STR intensity, neither STR position nor intensity are the major forcing and other large-scale modes of climate variability (possibly due to ENSO as the effects start to be felt in JJA) are important in these terminal nodes.

The years classified in each terminal node in JJA are noteworthy (Figure 11). The enhanced ‘winter-like’ group (Terminal Node 7) is rare throughout the instrumental record and has not occurred since 1981, making the last 28 years the longest stretch without an occurrence of this node. Also of interest is the disappearance of the wet Terminal Node 13, last observed in 1978. There is no other 30-year period, during the instrumental record, when this wet node has not occurred.

Figure 11.

SEA JJA precipitation time series with ENSO years marked by dots. Colour of the vertical bar indicates the terminal node the year has been classified into. The red and blue horizontal lines indicate the cut-off points for the original classification (dry, average and wet), giving an indication of misclassification when the horizontal and vertical classifications do not match.

The signal of the recent drought is found primarily in MAM (Murphy and Timbal, 2008), but there is evidence of a growing importance of the JJA rainfall deficit (Timbal, 2009). In the last two decades 7 years are classed as dry, with all of these dry years classified on the basis of high STR intensity into Terminal Node 2. However, this is not unprecedented as many periods in the past have large numbers of dry years classified on the basis of high STR intensity. Although the node purity in Terminal Node 2 is very high, three wet years appear in the dry terminal node. Interestingly, for over 90 years no wet years were found when STR intensity was high yet the three wet instances have occurred in the last 15 years – 1995, 2003 and 2005. This could be purely random as it is a very small number, but it is also interesting to note an increase in the occurrence of average rainfall years when STR intensity is high (Terminal Node 2) from 3 cases in the first half (1900–1955) of the record to 11 cases in the second half. Precipitation in SEA during these 3 years is much higher than other cases in the same terminal node (1.82 mm/day compared with 0.92 mm/day across the remainder of Terminal Node 2), which may be because of a warming of sea surface temperature (SST) surrounding Australia (Casey and Cornillon, 2001). The rainfall for these three wet years in Terminal Node 2 was analysed further. The spatial analysis reveals that the positive precipitation anomalies for the three wet years in SEA are widespread, with negative anomalies along the eastern coastal strip. The intensity of the STR is slightly lower on average in these wet cases (1023.8 hPa rather than 1024.1 hPa for Terminal Node 2); however, the STR is 2.4° further north during these 3 years (29.3°S compared with 31.7° for Terminal Node 2). An analysis of the moisture transport (not shown) points to the possibility that during these years, despite the northerly location of the STR, anomalous moisture flux originating from the Coral Sea (150–170°E, 5°S–25°N) reaches SEA from the north and may contribute to reduce easterly flow along the NSW coast (in agreement with the rainfall anomalies). This exploratory analysis is far from definitive and further research is required to confirm this hypothesis.

3.3. Interaction with ENSO

As ENSO events develop during JJA we may expect to see some interaction between ENSO and the STR in this season that is not evident in MAM. In the MAM analysis ENSO events are evenly spaced throughout the four terminal nodes as expected as there is little ENSO signal during this season (not shown). During JJA, however, El Niño events are much more tightly grouped with 18/25 events occurring in Terminal Node 2 (dots in Figure 11). Of the 18 El Niño events classified into this dry terminal node, all show average (3 years) or below average (15 years) rainfall. Interestingly, the remainder of the ENSO years classified into other terminal nodes are more evenly distributed with a counter-intuitive tendency towards being wet: one dry, three average and three wet El Niño years. This analysis suggests a strong interplay between ENSO and the STR in JJA. This supports the conclusions of Cai et al. (2011) that found a strong relationship between tropical variability (expressed through the Indian Ocean Dipole) and the STR in this season. Indeed ENSO variability contributes to STR variability as it can be seen that most of the El Niño years end up having above average STR intensity (Terminal Node 2). In addition, STR variability appears to be able to modulate El Niño impact on SEA rainfall: El Niño years classified in other nodes are a mix of wet, dry and normal years.

3.4. Vector wind analysis

Further analysis of the climatology of the terminal nodes is performed by looking at lower tropospheric wind in both MAM and JJA. The composite vector wind analysis for the years between 1948 and 2009 identified in each terminal node identified in the CART analysis can be seen in Figure 12 for MAM and Figure 13 for JJA.

Figure 12.

The 850 hPa vector wind MAM mean (top diagram) and composite anomalies (for the four terminal modes, below). Units are m/s. Computed from reanalyses from 1948 to 2009.

Figure 13.

The 850 hPa vector wind JJA mean (top diagram) and composite anomalies (for the four terminal modes, below). Units are m/s. Computed from reanalyses from 1948 to 2009.

Over the period 1948–2009, the mean flow over the study region is westerly airflow of between 2 and 6 m/s in MAM, picking up to 4–8 m/s in JJA, and intruding further over the Australian continent in JJA compared with MAM. In the dry Terminal Nodes of both MAM and JJA (both Terminal Node 2), the anomalous flow over the study region is south-easterly (MAM) to easterly (JJA) as part of a broad anti-cyclonic regime over the study region (with corresponding high rainfall in JJA along the east coast at the location of the strongest easterly anomalies as the GDR acts to uplift the moist oceanic air). The strongest anomalies in JJA. The anomalous wind flow in MAM's Terminal Node 7 and JJA's terminal Node 13 shows the southerly position of the STR, with easterly anomalies (i.e. weakening of the westerly flow) on the southern side of the STR. In the case of MAM, where Terminal mode 7 was linked to a ‘summer-like’ rainfall node, wind anomalies can be seen to originate from the tropics (Figure 12). Strong easterlies over north Queensland are reduced (1.5–2 m/s) and instead north-easterly anomalies affect SEA. The predominant westerly anomalies on the western and southern flank of the STR are strongly reduced (up to 2 m/s) and shifting south.

Finally, the anomalous vector wind composites for Terminal Nodes 12 and 13 in MAM (Figure 12) and Terminal Nodes 7 and 12 in JJA (Figure 13) show the different flavour of westerly mean flow being pushed further north and hence affecting SEA, and the strength of the anomalies and the location of the westerly maximum differ from case to case. The most obvious case is the wind anomalies for JJA's ‘winter-like’ Terminal Node 7; it displays the large westerly wind anomaly in excess of 2 m/s over inland Australia, suggesting a strong push of the westerlies further north. In the case of MAM's Terminal Node 12, wind anomalies are harder to reconcile with the rainfall anomaly maps than in the other previous terminal nodes. They show consistent westerly anomalies over SEA, yet a very limited rainfall signature apart from a small region on the Victorian coast and Tasmania.

4. Discussion and conclusions

This is a pioneering study introducing the use of CART to explore complex relationships within the climate system, and the large-scale influences of regional rainfall in SEA. The focus in this study is on the interplay between STR changes in both intensity and position, and SEA precipitation in MAM and JJA. CART analysis is used to complement results obtained with linear statistics.

The linear analysis confirms recent research showing the relationship between STR intensity and position with SEA rainfall in all seasons (Williams and Stone, 2009). The dominant role played by changes in intensity over position for SEA rainfall variability is highlighted, particularly in light of the current rainfall deficit in SEA, which is mainly a MAM phenomenon with an additional contribution in JJA (Timbal and Drosdowsky, 2013). Rainfall in both seasons displayed a stronger relationship with STR intensity than with position and followed a seasonal cycle with correlations reaching maximum values (spatial extent and magnitude) in JJA. However, a split in two 50-year epochs showed significant changes in the relationship between the STR and precipitation over time. Most notable was a change in the sign of the relationship with the position of the STR in MAM across SEA, which highlighted the importance of the mean position of the STR, particularly in a non-stationary climate. Correlation and regression analysis are traditionally used heavily to investigate relationships within the climate system (Williams and Stone, 2009, Timbal and Drosdowsky, 2013) so we turned to more complex nonlinear statistics to confirm and extend the current understanding. This is important because correlation analysis, particularly when sample size is limited to about hundred years, does not provide a robust understanding of more complex issues such as the interplay between position and intensity, i.e. linear regression does not indicate an additional role for changes in position beside the STR intensity (Timbal and Drosdowsky, 2013). The nonlinearity of the relationship hinted at by computing correlations over different epochs, and the complex interplay between changes in position and intensity was investigated further by using classification trees. Confirming our understanding of the system gained from the linear analysis, the CART analysis showed that intensity was the most important variable in both seasons for classifying dry years but also played a role in classifying wet years. Most importantly, the CART analysis was able to tease out the complexities of this interaction and how the two aspects of the STR (intensity and position) interact.

The CART analysis confirms that STR intensity is the most important variable for classifying dry conditions in both MAM and JJA; in both seasons, the first split identified dry terminal nodes with negative precipitation anomalies over SEA classified on the basis of high STR intensity. An analysis of vector wind anomaly composites for these nodes showed a reduction in the mean westerly flow across SEA. This is consistent with the spatial rainfall composite for these years, with negative precipitation anomalies across the entire region under the influence of frontal activity in both seasons. The years in each of the dry terminal nodes confirmed that the current rainfall deficit is linked to increasing STR intensity (Timbal, 2009). Indeed, in the last two decades a majority (11/13) of dry MAMs are classified in the dry terminal node (on the basis of high STR intensity) and while the rainfall deficit in JJA is not as severe, one can observe that all dry years in the last two decades are classified into the dry terminal node on the basis of high STR intensity.

But the real value of this approach is that the complexities of the STR position–rainfall relationship that were masked by the coarse linear analysis appear more clearly within the CART analysis. Indeed, the STR position is needed in both seasons to classify the remaining years. In MAM, position played a more prominent role when intensity is close to or below average, the secondary split in the CART analysis is based on position. This greater role for position in MAM is consistent with the large changes in position observed in this transition season. In JJA, position was only necessary for a third split between terminal nodes when intensity was neither significantly above or below average. Indeed in JJA, the relationship between rainfall in SEA and STR intensity appears more linear than in MAM with extreme values of intensity associated with the driest and wettest cases. In this season, the position of the STR is only playing an important role when intensity is near average. Overall, in both seasons, the fact that the role of the STR position depends on the intensity of the STR may explain why multiple linear regression statistics cannot find an additional role for the STR position independently of the STR intensity (Timbal and Drosdowsky, 2013). It is only once particular intensity conditions are met that position becomes important to determine the sign and spatial signature of precipitation anomalies over SEA.

One of the most revealing findings is the role that STR position has in MAM on the atmospheric circulation. The CART analysis revealed two wet regimes in MAM that are very different in their nature; a ‘summer-like’ rainfall node (Terminal Node 7) and a ‘winter-like’ rainfall node (Terminal Node 13). The ‘summer-like’ rainfall node shows that when intensity is low and the STR is in a southerly position wet conditions are evident over much of the east coast as tropical interactions bring moist air to southern Australia. This combines with a reduction of the mean westerly flow which pushes frontal activity south resulting in the negative precipitation anomaly on the Victorian coast, indicative of the southerly position of the STR. The highest rainfall anomalies are apparent on the eastern coast as the mean anti-cyclonic winds associated with the high-pressure cell bring a moist onshore airflow over the continent. The positioning of the STR at higher latitudes is consistent with a widening of the tropics and Hadley Circulation (Seidel et al., 2007), although the rare occurrence of this node does not enable us to draw very firm conclusions apart to note that it has occurred in the most recent decades. More importantly, the existence of this ‘summer-like’ node highlights that natural variability in the width of the tropical zone is associated with rainfall anomalies over eastern Australia as far south as SEA, despite the lack of relationship between rainfall in SEA and tropical modes of variability in MAM (Timbal et al., 2010). This ‘summer-like’ rainfall mode differs markedly from the ‘winter-like’ node (Terminal Node 13) and highlights the role of the STR position. When the STR is located in a northerly position (≤36.17°) the precipitation pattern is indicative of precipitation caused by cold fronts imbedded in a moist westerly airflow; positive anomalies confined to the SEA region with the largest values on the Alps and the west coast of Tasmania, suggesting an increased mean westerly flow.

Overall, the CART analysis shows that in JJA the effects of position are not well defined. This is consistent with the mean position of the STR, north of SEA during this season and fairly stable across the JJA months while in MAM it was noted to change dramatically from month to month from a position south of SEA to the north (Figure 1). Overall the CART analysis confirmed many known features identified using simple correlations. The increased node purity in all terminal nodes in JJA compared with that in MAM is consistent with the higher correlation between both STR intensity and position with rainfall in JJA. The threshold values upon which splits were based support the seasonal cycle, i.e. higher threshold values for STR intensity in JJA is a result of higher STR intensity during the JJA months compared with the MAM months. The precipitation patterns of JJA's dry and the wet terminal nodes (Terminal Nodes 2 and 7) are mirror images of each other (see also the vector wind analysis) and are associated with high and low STR intensity because of the linear nature of the relationship between STR intensity and rainfall in this season. Overall, the CART analysis confirms that the STR intensity–rainfall relationship is much more linear than in MAM and that the nonlinear relationship with the STR position play a minor role for rainfall as the STR is located further north. Terminal Nodes 12 and 13, however, are classified on the basis of STR position, and the composite analyses (precipitation and vector wind) of these nodes do not show many differences in SEA, although more sizeable differences are seen further north.

On the basis of the analysis provided here, one can conclude that while the position of the STR plays a clear role in MAM it is not the case in JJA and this is likely owing to the mean position of the STR in both seasons. In JJA, the mean position of the STR is to the north of the study region, and thus variation in the position has little effect on precipitation in SEA, while variation in STR intensity is still able to play an important role during this season. Variability in STR position has a more important association with precipitation in SEA in MAM, the transition season from a southerly to a northerly position, as the mean position of the STR is further south combined with its annual meridional movement.

A noted feature of the JJA CART analysis was a small number of wet years, three amongst the predominantly dry Terminal Node 2, which occurred in the most recent period (1995, 2003 and 2005). These three cases experienced much higher rainfall while the intensity of the STR remained high, although the mean position of the STR was further south in these 3 years compared to the full record. No wet cases were recorded during the first 94 years of the instrumental record yet three wet cases were recorded in the last 15 years. The anomalous vector wind analysis of these years showed that a moist northerly airflow dominated SEA. Analysis of the Coral Sea SST showed that all 3 years had positive SST anomalies in this region; in addition 2003 and 2005 were in the top 10th percentile. This shows that in rare occasions, despite high STR intensity, a small shift southward of the STR position, in combination with warm SST anomalies in the Coral Sea region, may affect precipitation in SEA. It was suggested that perhaps these occurrences have only recently appeared in the record because of increased regional SST (Casey and Cornillon, 2001). Although logical, this hypothesis would require more in-depth analysis to be demonstrated. Less controversial were the results showing interplay between El Niño events and STR intensity in JJA, i.e. El Niño events predominantly being observed in conjunction with above-average STR intensity and resulting predominantly in dry years, while other El Niño years not associated with above average STR intensity were predominantly average or wet years. This suggests that both ENSO variability impacts STR variability, and that independent STR variability can also modulate the impact of ENSO on SEA rainfall.

It is clear that in recent years STR intensity has been important for precipitation in SEA (Timbal and Drosdowsky, 2013) but that the role of position should not be underestimated, particularly in MAM, despite the lack of additional effect reported in that same study (Timbal and Drosdowsky, 2013). Indeed it has been shown here that MAM precipitation variability in SEA cannot be described solely by STR intensity and that wet years in particular require the role of position to be understood. The mean STR intensity in MAM is 1020.2 hPa, and we see in the classification tree that over the instrumental record when the intensity is ≥1021 hPa conditions are dry 60% of the time. The implications for this relationship in a climate change regime are interesting; while the magnitude is unclear at this stage, changes to the Hadley Circulation under a climate change regime are likely (Seidel et al., 2007). Owing to its geographic position, SEA is highly sensitive to changes in the STR position. Consequently, the subtle effect position has precipitation in SEA, while not as important as the impact of rising intensity, should not be discarded.


This study was supported, in part, by the South Eastern Australian Climate Initiative. We thank Carolyn Whan and Clive Hilliker for assistance with graphics. We are grateful to two anonymous reviewers whose suggestions improved the quality of this article.