Springtime rainfall changes in Australia related to projected changes in large-scale modes of variability

3.1. Model consensus on driver amplitude changes

In this section, we identify the models that show the largest consensus on the sign of amplitude change in ENSO, IOD and SAM. Fig. 1 shows the list of models exhibiting a change of >0.5 s.d. of all projected amplitude changes: increase (red) or decrease (blue) in N3.4, DMI and SAM indices. The threshold of 0.5 s.d. is chosen so that only models displaying relatively large changes are selected for each group (increase or decrease) while still maintaining a sufficient sample size for significance testing. Additional analysis using a 1.0 s.d. threshold was performed, with fewer models being selected, but showing very similar associated rainfall changes (not shown).

Fig. 1.

Models showing projected MMEM increases (red) and decreases (blue) in Niño3.4 (top row), DMI (middle row), and SAM (bottom row) variability, comparing strengths of 2035–2099 (late 21st Century) events relative to 1950–2014 (late 20th Century) events. Increases and decreases are measured for each driver separately. Only models with an amplitude change of over 0.5 s.d. (across all projected changes) are coloured. Models are listed in alphabetical order.

Fig. 1 shows that all 33 models show a large change in at least one driver. Nineteen models show a large increase in overall N3.4 amplitude (‘ENSO N3.4’ row), with five showing a large decrease. Nine models show little to no change. Fifteen models exhibit a large decrease in overall DMI amplitude (‘IOD DMI’ row), with 6 models showing a large increase, and 12 models showing small or no change. For SAM, 22 models exhibit a large amplitude increase. Only five models exhibit a large amplitude decrease, and six models exhibit little or no change. The most common changes therefore correspond to an increase in ENSO and SAM variability, and a decrease in IOD variability. These changes are estimated to be statistically significant at the 99, 99.9 and 92% levels respectively according to a two-tailed binomial test.

To gauge if the interrelationship between these changes to ENSO, IOD and SAM are linked, we try to identify if the changes are correlated. If we assign +1 to an increase, −1 to a decrease and 0 to little or no change in each model (focusing here only on the direction of change, not the magnitude), then we can calculate the correlation coefficient between the sign of the driver changes. The correlation between the sign of changes in ENSO and the IOD is 0.37. The correlation between the sign of changes in ENSO and SAM is −0.01. The correlation between the sign of changes in the IOD and SAM is 0.14. This indicates that the changes in SAM occur independently of the other two drivers, whereas the changes in the IOD are weakly (and positively) correlated with changes to ENSO.

These results help to explain why only 5 of the 33 models (15%) exhibit an increase in the variability of both ENSO and SAM, accompanied by a decrease in IOD variability. So, although the most common responses are increases in ENSO SST variability (in 58% of models), an increase in SAM variability (in 67% of models) and a decrease in IOD variability (in 45% of models), only 15% of the models exhibit all three of these changes.

For ease of presentation, for the rest of the study, we focus on the most frequently evident change in the models. To do this, we restrict attention to the 19 models that exhibit an increase in ENSO variability when discussing changes in ENSO teleconnections, the 15 models that exhibit a decrease in IOD variability when discussing changes in IOD teleconnections, and the 22 models that exhibit an increase in SAM variability when discussing SAM teleconnection changes. We refer to these groups of models as the ‘representative models’ in the rest of this study. The rainfall changes evident in the remaining ‘minority’ groups of models are shown in Supplementary Fig. S3–S8.

3.2. Overall teleconnection changes

Over the 1950–2005 period, observed springtime rainfall in large parts of central and eastern Australia was significantly correlated with Niño3.4 and the DMI, and rainfall in southern and eastern Australia was correlated with the SAM index (not shown). This correlation is reasonably well simulated in the CMIP6 multi-model mean for ENSO and IOD, but less so for SAM (Chung et al. 2023). Although Chung et al. (2023) used only the first ensemble member for each model, we find similar results using all available ensemble members from the same models.

Fig. 2 shows the SON teleconnection pattern (i.e. the linear regression coefficient between Niño3.4, DMI and SAM indices, and rainfall) for Historical and SSP3-7.0 scenarios. For each driver, we use the subset of ‘representative models’ identified in the previous section. Fig. 2a shows a significant relationship between Niño3.4 and rainfall over north-eastern, central and south-eastern Australia in the historical late 20th Century period. Here, significance is measured as any grid point in which >70% of models exhibit a regression coefficient that is significant at the 90% level using a t-test. Under SSP3-7.0, the regions of significance extend over northern and eastern Australia, strengthening slightly in the north and in southern Victoria (Fig. 2d). Fig. 2g shows the change in teleconnections (late 21st Century minus late 20th Century). This is broadly consistent with McGregor et al. (2022), who analysed the global projected change in DJF ENSO teleconnections, finding a strengthening in the ENSO teleconnections to north-eastern Australian rainfall.

Fig. 2.

Multi-model mean of linear regression slopes between SON mean rainfall and (a) SON Niño3.4, (b) SON DMI, (c) SON SAM indices for 1950–2014 (late 20th Century) using the subset of ‘representative models’ for each driver. (d–f) The same, but for 2035–2099 (late 21st Century) under the SSP3-7.0 scenario. (g–i) Late 21st Century minus late 20th Century regression slopes. To ensure each model is weighted equally, the multi-ensemble mean is used for models with more than one ensemble member. In (a–f), maps are stippled where more than 70% of models show a regression coefficient that is significant at the 90% level according to a t-test.

The IOD teleconnection also is projected to strengthen from late 20th to late 21st Centuries, with a band of significant regression coefficients extending from north-western to south-eastern Australia in the late 20th Century (Fig. 2b) expanding to cover a larger fraction of central and northern Australia in late 21st Century (Fig. 2e, h). It is interesting to note that IOD teleconnections are projected to strengthen even in this subset of models that project an overall weakening of the DMI. This may be linked to several contributing factors. Firstly, Wang G et al. (2024) showed that models project increased SST variability in the tropical Indian Ocean that is not fully captured by the DMI owing to asymmetric responses in the eastern and western regions of the Indian Ocean. Secondly, rainfall during IOD years is closely related to the strengthening ENSO teleconnections, as many IOD and ENSO events co-occur. Thirdly, it is likely that in a warmer world, SST variability in the tropics can inducea stronger rainfall response (e.g. Pendergrass 2020). In Section 3.4, we take a closer look at the asymmetric rainfall response during pIOD and nIOD phases.

For SAM, as noted in Chung et al. (2023), there is generally a poorer representation of teleconnections and a larger amount of inter-model spread. In this subset of models, in late 20th Century, the only significant teleconnections are found in western Tasmania (Fig. 2c). In late 21st Century, this teleconnection strengthens and extends to the south-eastern coast of Australia (Fig. 2f, i).

The differences between late 21st and late 20th Centuries regression slopes are shown in Fig. 2g–i, highlighting the regions in which the relationship strengthens significantly. For ENSO and IOD, these are largest in northern Australia. The representative models therefore indicate that ENSO, IOD and SAM teleconnections are likely to strengthen under SSP3-7.0 in many regions, particularly over central and eastern Australia. This is true even in the case of the IOD, which is projected to weaken.

We now consider two questions:

Here, we present the MMEM changes to rainfall totals and anomalies averaged over ENSO, IOD and SAM phases. For ease of presentation, we show only the ‘representative’ model MMEM for each driver discussed in Section 3. For reference, MMEM SON rainfall changes for all years are shown in Supplementary Fig. S1a, and late 20th Century anomalies for each phase of ENSO, IOD and SAM are shown in Supplementary Fig. S1b–g.

3.3. Stronger ENSO events

From Section 3.1, we showed that the MMEM of the subset of ‘representative models’ for ENSO events indicates stronger ENSO activity, or, more precisely, a larger amplitude in N3.4 SST swings. The MMEM rainfall projections for these 19 models are shown in Fig. 3, separated into El Niño (EN) and La Niña (LN) years. We present the rainfall projections in two ways: the mean-state change (Fig. 3a, c), and the change in anomalies during EN and LN years (Fig. 3b, d). The mean-state change is simply the late 21st–late 20th Century difference in rainfall averaged over all years in a particular phase, for example during EN or LN years. The change in anomalies is the late 21st–late 20th Century difference in EN and LN anomalies. For example, EN anomalies in the late 21st Century period would be the rainfall during EN years minus the mean rainfall for all years in late 21st Century. EN anomalies in the late 20th Century period would be the same, but relative to the late 20th Century climatology. The change in anomalies is then the difference in these two quantities. For reference, late 20th Century anomalies for EN and LN years are shown in Supplementary Fig. S1b, c and late 20th Century correlations between rainfall and N3.4 are shown in Supplementary Fig. S2a, b.

Fig. 3.

(a) SSP3-7.0 (late 21st Century) – historical (late 20th Century) mean-state change in total SON rainfall, averaged over El Niño years only. (b) SSP3-7.0 (late 21st Century) – historical (late 20th Century) change in SON rainfall anomalies, averaged over El Niño years only. (c, d) The same, but for La Niña years. Maps show the multi-model mean of the 19 ‘representative’ models that project stronger ENSO variability. Maps are stippled where over 70% of the models agree on the sign of the change.

The ‘representative’ models project a significant mean-state drying over south-west Western Australia (SWWA) and southern Victoria during SON in both EN and LN phases (Fig. 3a, c). During EN years, there is also a significant mean-state drying over most of southern and eastern Australia (Fig. 3a). EN-related rainfall deficits are projected to intensify over south-eastern Australia and over the Kimberley region in north-west Australia (Fig. 3b). During LN years, the mean-state drying over SWWA and southern Victoria is accompanied by a rainfall increase around the Kimberley region (Fig. 3c). LN-related anomalies are projected to become wetter over much of southern Australia and parts of the Northern Territory (Fig. 3d).

The ‘minority’ models corresponding to a decrease in ENSO variability (five models), or little to no change (nine models) are shown in Supplementary Fig. S3–S4. The significant mean-state drying over SWWA and southern Victoria appears in all groups of models, though other regional mean-state changes differ between groups. The EN- and LN-related anomalies also differ, with the ‘weaker ENSO’ group generally exhibiting overall opposite sign changes from those shown in Fig. 3, as expected. The ‘no change’ group exhibits wetter LN-related anomalies but little change in EN-related anomalies, reflecting a general strengthening of teleconnections especially during LN years.

In the ‘representative’ models corresponding to stronger ENSO variability, the overall drying during EN and wetter conditions during LN years are consistent with a stronger driving ENSO signal from the Pacific, and consistent with the strengthened teleconnections shown in Fig. 2a, d. We note also that the mean-state drying over SWWA and southern Victoria is a robust trend that occurs regardless of driver change or phase, and is likely dominated by the climate change signal. SWWA is not a region that is heavily influenced by ENSO (Fig. 2a), so it is unsurprising that the mean-state drying over this region occurs in all models. However, southern Victorian rainfall is significantly correlated with ENSO, so it is notable that the mean-state drying occurs in all models, even under strengthened teleconnections to ENSO in which LN-induced rainfall increases; see discussion below.

3.4. Weaker IOD events

In contrast to ENSO, for the IOD, the 15 ‘representative models’ project weaker events. In these projections, both pIOD and nIOD years exhibit mean-state drying over much of western and eastern Australia, and a slight rainfall increase over the Kimberley region (Fig. 4a, c). pIOD years exhibit more widespread mean-state drying over eastern Australia (Fig. 4a). From Fig. 4a, c and Supplementary Fig. S5–S6, we can see again that the mean-state drying over SWWA and southern Victoria arises regardless of driver change or phase, noting that SWWA is also a region that is not typically heavily influenced by the IOD. The mean-state rainfall increase over the Kimberley region during pIOD years also occurs in all model groups.

Fig. 4.

(a) SSP3-7.0 (late 21st Century) – historical (late 20th Century) mean-state change in total SON rainfall, averaged over positive IOD years only. (b) SSP3-7.0 (late 21st Century) – historical (late 20th Century) change in SON rainfall anomalies, averaged over positive IOD years only. (c, d) The same, but for negative IOD years. Maps show the multi-model mean of the 15 ‘representative’ models that project weaker IOD variability. Maps are stippled where over 70% of the models agree on the sign of the change.

Late 20th Century rainfall anomalies and rainfall–DMI correlation for pIOD and nIOD years are shown in Supplementary Fig. S1d, e and S2c, d respectively. In a future with weaker IOD events, we would typically expect IOD-related rainfall anomalies to weaken, i.e. that pIOD anomalies would be less dry and nIOD anomalies would be less wet. Although this is certainly true of pIOD anomalies, nIOD anomaly changes in fact strengthen. The blue shading in Fig. 4b indicates that pIOD rainfall anomalies, which are typically dry, become less dry over parts of central and south-eastern Australia. However, nIOD-related anomalies become wetter (Fig. 4d), which reflects an asymmetry in the response to IOD change. It is also important to recall that the IOD and ENSO are strongly correlated in observations and models, so nIOD years often co-occur with LN years. The strengthening of wet anomalies during nIOD years despite the weakening of IOD variability may be related to the strengthened teleconnections to LN, given that many LN and nIOD years co-occur.

The ‘minority’ models corresponding to a strengthening IOD (6 models) and little to no change (12 models) show similar results for nIOD-related anomalies, which are increased rainfall in the south-east (Supplementary Fig. S5d, S6d). For pIOD-related anomalies, the ‘stronger IOD’ models show increased drying over the south-east and less drying in the west (Supplementary Fig. S5c), with similar but weaker patterns for the ‘no change’ models (Supplementary Fig. S6c). For pIOD years, changes are broadly of opposite sign to those exhibited by the ‘representative’ models.

3.5. Stronger SAM variability

Although there is limited model skill in capturing SAM teleconnections to Australia, SAM variability is projected to strengthen in 22 models, which is the most robust change out of all three drivers (Fig. 1). Similarly to ENSO and IOD, the significant mean-state changes for both positive SAM (pSAM) and negative SAM (nSAM) regions correspond to drying over SWWA, southern Victoria and southern Queensland (Fig. 5a, c). For nSAM years, the drying is more widespread across eastern Australia and there is a mean-state rainfall increase over the Cape York region (Fig. 5c). Neither pSAM- nor nSAM-related anomalies exhibit any widespread significant change (Fig. 5b, d), with only increased drying over the Kimberley region in nSAM years being stippled in Fig. 5d.

Fig. 5.

(a) SSP3-7.0 (late 21st Century) – Historical (L20C) mean-state change in total SON rainfall, averaged over positive SAM years only. (b) SSP3-7.0 (late 21st Century) – Historical (L20C) change in SON rainfall anomalies, averaged over positive SAM years only. (c, d) The same, but for negative SAM years. Maps show the multi-model mean of the 22 models that fall under the ‘most likely’ SAM-change pathway (stronger SAM variability in SON). Maps are stippled where over 70% of the models agree on the sign of the change.

Only five models project a weakening of SAM and four project little to no change. The mean-state and anomaly changes for these ‘minority’ groups of models exhibit varying regional responses (Supplementary Fig. S7–S8); however, the mean-state drying over SWWA again remains a robust response in all model groups. The ‘weaker’ SAM models generally show wetter anomalies in both phases of SAM, whereas the ‘no change’ SAM models show drier anomalies over the south-east in both phases of SAM.

For reference, late 20th Century rainfall anomalies in pSAM and nSAM years are shown in Supplementary Fig. S1f, g, with the late 20th Century correlations between rainfall and SAM index shown in Supplementary Fig. S2e, f. Given the relatively weak change in SAM teleconnections in the ‘representative’ models (Fig. 5b, d), it is therefore unclear if or how a strengthening of SAM variability during SON might affect rainfall variability. It is also worth noting that the SAM–ENSO relationship, which modulates the teleconnection of SAM with Southern Hemispheric rainfall, is not well captured in models (Lim et al. 2016). However, the mean-state changes over southern Australia clearly reflect the dominance of the mean-state climate change signal over any forced changes in internal variability.

Previous studies have shown that models project an overall drying over Australia during winter and spring (Delage and Power 2020; CSIRO and Bureau of Meteorology 2024), particularly over SWWA (Delage and Power 2020; McKay et al. 2023; Rauniyar et al. 2023) and southern Victoria (Rauniyar and Power 2023). This is evident in Fig. 3–5, and Supplementary Fig. S3–S8, which show that regardless of the direction of projected driver changes, or phases, SWWA and southern Victoria exhibit a significant drying signal. These projections indicate that the climate-change signal in these regions dominates the mean-state change in springtime rainfall. That is, regardless of how the variability of ENSO, IOD or SAM might change, over 70% of models project drying in the late 21st Century period in these regions. This is consistent with the CMIP5 results obtained by Delage and Power (2020) in their examination of the combined impact of global warming and ENSO on winter and spring rainfall over Australia.

3.6. Concurrent ENSO and IOD events

Given the impacts of EN + pIOD and LN + nIOD events on Australian rainfall are amplified when these phases co-occur, which they often do (e.g. Ummenhofer et al. 2011; Holgate et al. 2022), it is of interest to present similar rainfall projections for concurrent events. However, the projected strengthening in ENSO events and weakening in IOD events means it is not immediately clear how modelled rainfall might respond to future concurrent events. Therefore, in this section, we turn our attention to changes to the frequency (i.e. number) of concurrent events rather than the amplitude of the indices. Chung et al. (2024) compared the frequency of concurrent events in late 20th and 21st Centuries in each model, finding that there were almost no statistically significant changes in occurrences of concurrent events under SSP3-7.0. Only the LN + nIOD + nSAM combination exhibited a statistically significant increase.

As there is no clear majority in the direction of change, we present here rainfall changes associated with both EN + pIOD and LN + nIOD increases and decreases (combining both phases of SAM). We present MMEM mean-state changes and anomaly changes for four subgroups of models, i.e. those that project an increase or decrease in EN + pIOD and an increase or decrease LN + nIOD events. We select only models that exhibit a frequency change of >0.5 s.d. across all models (Fig. 6a). In doing this, there are 15 out of 33 models that show increases, and 10 that show decreases in EN + pIOD events (with eight models showing little to no change in frequency). By contrast, there are eight models each showing an increase in LN + nIOD and seven showing a decrease in LN + nIOD events (and 18 models showing little to no change). Note that in Fig. 6a, models with little to no change in either EN + pIOD or LN + nIOD are not listed.

Fig. 6.

Top row (a): models showing projected MMEM increases (red) and decreases (blue) in the frequency of EN + pIOD events and LN + nIOD events. Only models with an amplitude change of over 0.5 s.d. across all models are coloured. Models are listed in alphabetical order. Middle row (b–e): concurrent EN + pIOD events. (b) Projected mean-state change in total rainfall, averaged over EN + pIOD years, for models showing a large frequency increase in EN + pIOD events. (c) Projected change in rainfall anomalies, averaged over EN + pIOD years, for models showing a large frequency increase in EN + pIOD events. (d) The same as (a) but for models showing a decrease in EN + pIOD events. (e) The same as (c) but for models showing a decrease in EN + pIOD events. Bottom row (f–i): the same, but for LN + nIOD events and for LN years only. Projected changes are shown for SSP3-7.0 (late 21st Century) – historical (late 20th Century) scenarios, for SON rainfall. Maps are stippled where over 70% of the models agree on the sign of the change.

Fig. 6b–i shows the MMEM mean-state change (Fig. 6b, d, f, h) and anomaly changes (Fig. 6c, e, g, i) during concurrent EN + pIOD and LN + nIOD years associated with increases and decreases of EN + pIOD and LN + nIOD events. In the models that project an increase in EN + pIOD events, EN + pIOD years exhibit a mean-state drying over much of southern, central and eastern Australia, largest around SWWA and south-eastern Australia (Fig. 6b). (EN + pIOD)-related anomalies for these models show intensified drying over central and south-eastern Australia (Fig. 6c). This intensified drying over the south-east is consistent with what is expected from an increase in EN + pIOD events. The models that project a decrease in EN + pIOD also exhibit significant, though weaker, mean-state drying over regions covering SWWA, southern Victoria and south-eastern Queensland (Fig. 6d). The (EN + pIOD)-related anomalies for these models show wetter, or less dry, anomalies (indicated by blue shading) over parts of western and eastern Australia (Fig. 6e). The weaker anomalies seen here are broadly consistent with a weaker model response to fewer EN + pIOD events.

For LN + nIOD years, an increase in LN + nIOD events corresponds to a wetter mean state over northern Australia but still drying over SWWA (Fig. 6f). The LN-related anomalies for these models show an intensification of the wet anomalies over most of the country, particularly over the south-eastern and north-western regions (Fig. 6g). Again, this is consistent with what is expected from an increase in the rain-promoting phases of the drivers. The models showing a decrease in LN + nIOD events show mean-state drying over much of WA, with mixed changes elsewhere (Fig. 6h). (LN + pIOD)-related anomalies in these models still show an intensification of wet anomalies over much of eastern Australia, but drier anomalies over WA (Fig. 6i). Although there are fewer driving LN + nIOD events, the increase in wet anomalies over eastern Australias indicates that the future events that do occur have a stronger impact, which is consistent with the increase in teleconnection strength in these regions associated with stronger ENSO variability.

Although there is a clear lack of consensus on future changes to the frequency of concurrent ENSO and IOD phases, these results provide an indication of what the current generation of models project for either sign of change. More work is required to quantify the amplitude changes of these events as well as a detailed analysis of asymmetry in the responses to positive and negative phases.

3.7. Consecutive ENSO and IOD events

In recent years, consecutive (back-to-back) events such as the double ENs and pIODs of 2018–2019 (e.g. Wang and Cai 2020) and triple LNs and nIODs of 2020–2022 have also had major impacts on SON Australian rainfall (e.g. Huang et al. 2024). In this section, we identify all occurrences of two or more back-to-back EN, LN, pIOD and nIOD events and compare the projected rainfall changes for these events. For each back-to-back event in each model, the mean rainfall, or rainfall anomalies, averaged over the 2 years is calculated. This is then averaged over the whole period.

Mean-state changes and changes to anomalies are shown in Fig. 7 for consecutive EN (a, b), LN (c, d), pIOD (e, f) and nIOD (g, h) events. Chung et al. (2024) found no significant changes in any of these events under SSP3-7.0, so in this section, we present the MMEM of all 33 models. The mean-state changes for all four events, once again, display the strongest drying over SWWA and southern Victoria. For consecutive EN years, there is increased mean-state drying over much of south-west and eastern Australia (Fig. 7a). A weak response in the rainfall anomalies is seen, with some model consensus on increased drying in small parts of Victoria and the Kimberley region (Fig. 7b). Consecutive LN years show a mean-state increase over the Kimberley region and Cape York coast in addition to the drying in the south mentioned above (Fig. 7c). There is also a weak intensification of wet anomalies over parts of the south-east and north of the country (Fig. 7d).

Fig. 7.

Top row: (a) projected mean-state change in total rainfall, averaged over consecutive EN years, for all models. (b) Projected change in rainfall anomalies, averaged over consecutive years, for all models. (c, d) The same as (a, b) but for consecutive LN events. Bottom row: (e–h) the same as (a–d) but for consecutive pIOD (e, f), and nIOD (g, h) events. Projected changes are shown for SSP3-7.0 (late 21st Century) – historical (late 20th Century) scenarios, for SON rainfall. Maps are stippled where over 70% of the models agree on the sign of the change.

Consecutive pIOD and nIOD events display largely similar, albeit slightly weaker, responses to the EN and LN cases respectively. Consecutive pIOD years show mean-state drying over the south and eastern coast (Fig. 7e), with no model consensus anywhere on anomaly changes (Fig. 7f). Consecutive nIOD events show the mean-state drying to the south and increased rainfall over the far north (Fig. 7g), with models only agreeing on larger wet anomalies over a small part of Victoria (Fig. 7h).

These results suggest that the overall drying trend in SWWA, southern Victoria and Tasmania are a response to the changes in the mean state. Consecutive EN years would reinforce these dry conditions in future, whereas consecutive LN years would somewhat alleviate the mean-state drying. South-eastern Australia may therefore need to rely more on isolated and consecutive LN events in water management planning activities.