Introduction

The intensity and frequency of extreme precipitation events are widely expected to increase as a result of climate change1,2,3,4, which will have significant implications for future flooding5,6. Globally, observations show an increase in annual maximum daily precipitation between 5.9% and 7.7% per degree of warming7. Even larger increases have been reported for the 99th (11% K−1) and 99.97th (13% K−1) percentile precipitation8. Observed changes are a result of thermodynamic and dynamic processes, both of which are influenced by a warming atmosphere9. Thermodynamic processes such as the Clausius-Clapeyron (CC) relationship, which increases the water-holding capacity of the atmosphere by 6–7% per degree of warming9, lead to relatively spatially homogeneous changes in precipitation intensity on the order of 4–8% per degree of warming10. Daily precipitation extremes have been shown to be intensifying at a rate approximately equal to the CC relationship11. Dynamic processes substantially modify the regional response to climate warming, through changes to the frequency and intensity of synoptic and subsynoptic phenomena, including tropical and extratropical cyclones10. Increased warming also increases atmospheric stability, which weakens circulation, potentially reducing the intensity of precipitation extremes. On the other hand, latent heat release results in atmospheric instability that can strengthen storms and precipitation extremes, with the most extreme events seeing the largest increases12,13. The overall impact of climate change on precipitation intensity is therefore highly complex and regionally dependent.

Climate models are currently the best physically based approaches available to understand future precipitation processes, characteristics, and impacts. Global Climate Models (GCMs) from the latest Coupled Model Intercomparison Project Phase 6 (CMIP6) have shown increases to precipitation extremes in-line with the CC relationship, with smaller increases noted for more common events compared to rarer events13,14,15,16. Abdelmoaty & Papalexiou17 showed historic 1 in 100 Annual Exceedance Probability (AEP) precipitation events would occur approximately every 50 to 70 years for the Northern and Southern Hemispheres, respectively, by mid-century (2035–2067). Using 25 GCMs, Gründemann et al.15 found the intensity of 1 in 100 AEP precipitation events increased by between 13.5% and 38.3%, depending on the emissions scenario, with smaller increases for more frequent events. An ensemble of CMIP6 GCMs projected increases to 1 in 10 and 1 in 50 AEP daily and 5-daily precipitation at rates approximately equal to the CC relationship (~7% K−1)16.

GCMs are, however, applied at relatively coarse resolutions (~150 km) and have difficulty adequately representing precipitation patterns over complex terrain and resolving extreme precipitation processes18. These coarse resolution models are not always suited to providing reliable regionally specific information required to support adaptation and decision-making at regional scales. To overcome these issues, regional climate models (RCMs) dynamically downscale GCMs to a finer resolution (typically 10–50 km) in order to better represent small-scale features and processes. RCMs have been shown to have improved skill in representing patterns of local precipitation and the impacts of topography, coasts, and land use changes compared to GCMs19,20,21,22 and may also be more skilful in simulating extremes20,23,24.

In Australia, most studies to date using RCMs have focused on “common” extremes, which typically have probabilistic return periods of a year or less such as indices defined by the Expert Team on Climate Change Detection and Indices25, the annual maxima (AM, RX1day), or percentiles (e.g., 95th, 99th, and 99.7th percentile)26,27,28. While common extreme events are important for water resources, agriculture and ecosystems, most flood impacts are associated with “rare” extremes, which have multi-year or multi-decade return periods and are less well understood15. These rare extremes are also used to inform floodplain management and infrastructure design throughout Australia, and as such, any increase in the magnitude of these events would have broad implications for future infrastructure and planning. Herold et al.29 projected 1 in 20 AEP daily precipitation would occur 1–2 times more frequently in capital cities across southeast Australia using projections from 4 downscaled CMIP3 GCMs, though there was low model agreement for these changes. Mantegna et al.30 showed daily intensities could increase by over 15% K−1 for the most extreme events (1 in 100 AEP) over Tasmania using 5 downscaled GCMs, with smaller increases for more common events. Similarly, Wasko et al.31 found larger increases for more extreme daily events (up to the 1 in 50 AEP) compared to less extreme events across Australia using four statistically downscaled GCMs, with larger increases for the north and smaller increases for the southwest.

There are significant uncertainties associated with projections of precipitation from climate models32, which necessitates large ensembles to better account for this uncertainty, especially when evaluating extremes33. However, studies on rare extremes to date have largely adopted limited ensemble sizes and used earlier climate projections, which may underestimate the uncertainty. There is a need to determine how climate change influences precipitation extremes using a large ensemble of the latest downscaled CMIP6 projections, particularly for the rarest events.

The focus of this paper is therefore to examine the impacts of climate change on precipitation extremes using a large ensemble of downscaled CORDEX-CMIP6 projections. For this purpose, we use downscaled ensembles from 4 modelling groups under a high emissions Shared Socioeconomic Pathway (SSP370), as only this scenario and SSP126 were shared by all modelling groups. SSP370 is a high emissions scenario with a relatively monotonic trend in radiative forcing, which allows for a more sensitive response at the end of the century. By contrast, SSP126 has a relatively low change in forcing and is parabolic across time. This large ensemble forms the core of Australia’s new national projections, contributing to the CORDEX-CMIP6 Australasia domain20,34,35,36, and underpins Australia’s climate services and adaptation planning. This large ensemble consists of 19 different host GCMs downscaled using 3 distinct regional climate models in 5 configurations, for a total of 39 different downscaled simulations (Table 1; more details in methods section). This represents the largest ensemble of high-resolution climate projections used to assess changes to precipitation in Australia, allowing for a more robust assessment of the uncertainty and the spatial variability of likely changes. The inclusion of multiple different downscaled ensembles allows for an assessment of the uncertainty introduced from the downscaling approach, which has not previously been quantified for rare precipitation extremes. We focus on greater capital cities as this is where two-thirds of the Australian population lives and thus where the impacts are likely to be most pronounced. The aims of this paper are threefold:

  1. i)

    To assess the performance of the CMIP6 GCMs and different RCM ensembles in simulating mean and common extreme precipitation (i.e., 99th and 99.7th percentiles).

  2. ii)

    To investigate changes and uncertainty in mean, common extreme (99th and 99.7th percentiles), and rare extreme precipitation (1 in 10, 50, and 100 AEP) per degree of global warming based on a large ensemble of regional models.

  3. iii)

    To investigate how extreme precipitation varies spatially across the continent and identify regions with more pronounced extreme precipitation increases.

Table 1 Details of the different ensembles of climate considered in this study, consisting of 15 simulations from QldFCP-2 (oc denotes when downscaling involved ocean coupling), 7 simulations from CCAM-ACS, 7 simulations from BARPA-ACS, and 10 simulations from NARCliM2.0
Full size table

Results

Evaluation of simulated daily mean and common extreme precipitation

We assess the performance of the host GCMs and the four downscaled ensembles in simulating daily mean and common extreme precipitation across Australia by comparing them to observation data. The ensemble of CMIP6 GCMs performed reasonably well in representing mean and common extreme precipitation across Australia, with lower average errors over Australia than some of the downscaled ensembles (Fig. 1). The performance of the CMIP6 ensemble was, however, lacking over coastal and mountainous regions, with a consistent dry bias, particularly for the common extremes (p99 and p99.7). This is evident when evaluating biases for the greater capital city regions, which are all situated near the coast and/or exhibit complex terrain. In these regions, where the majority of the population resides, the CMIP6 GCMs were shown to perform significantly worse in representing common extremes compared to the downscaled ensembles, though performance was comparable when evaluating mean precipitation (Fig. 2). Of the downscaled ensembles, NARCliM2.0 and QldFCP-2 had notable dry biases over northern and western Australia, with smaller biases over southeast Australia (Fig. 1). By contrast, CCAM-ACS was noted to have an opposing wet bias that was greatest in the interior, including for southeast Australia. BARPA-ACS was the best-performing ensemble over Australia as a whole for both mean and common extreme precipitation, with low mean absolute percentage error (MAPE) and root mean square error (RMSE) values.

Fig. 1: Ensemble average bias of daily mean, 99th percentile, and 99.7th percentile precipitation (model–AGCD) for the CMIP6 host models, QldFCP-2, CCAM-ACS, BARPA-ACS, and NARCliM2.0 ensembles over the 1981–2020 time period.
figure 1

Model comparisons are performed on a common 10 km grid, and highlighted values denote the MAPE (black) and RMSE (grey), stippling shows where the signal-to-noise ratio >1.0. Areas where precipitation is poor in observations are masked out.

Full size image
Fig. 2
figure 2

Annual averages of daily mean, 99th percentile, and 99.7th percentile precipitation from observations and ensemble mean bias (model–AGCD) over the 1981–2020 time period for the greater capital city regions.

Full size image

Within the greater capital city regions, all downscaled ensembles reported lower biases for p99.7 compared to mean precipitation and p99, except for CCAM-ACS, which had lower biases for mean precipitation (Table S1 to S3). For mean precipitation, CCAM-ACS had the lowest bias of any ensemble in the capital cities (Table S1). QldFCP-2 and NARCliM2.0 both showed a clear dry bias over most capital cities for mean precipitation, whereas biases for CCAM-ACS and BARPA-ACS varied according to the city. For p99 and p99.7, the performance of all downscaled ensembles varied according to location (Fig. 1). For p99.7, CCAM-ACS tended to overestimate precipitation intensity in capital cities, whereas NARCliM2.0 tended to underestimate intensity, though biases for NARCliM2.0 were relatively small. In contrast to the mean precipitation, QldFCP-2 also tended to have low biases for p99.7, except for in Perth and Hobart, where the models showed a dry bias. BARPA-ACS also had good agreement except for in Brisbane and Adelaide, where a wet bias was noted (Table S3). These biases are, however, in almost all cases, an improvement when compared to the CMIP6 host models for the capital cities, suggesting that the downscaled models from these ensembles may be better suited to represent precipitation extremes in these regions compared to the host models.

Projected changes to mean, common and rare extreme precipitation

Considerable spatial variability was evident in the projections of future mean and common extreme precipitation over Australia (Fig. 3). Spatial patterns of change varied according to the ensemble applied, with a consensus towards wetting or drying evident in a few regions only. A clear drying pattern for mean precipitation was evident in southwest Western Australia and southern Victoria and Tasmania to a lesser extent. This, however, did not necessarily lead to clear decreases in common extreme precipitation over these regions. In other regions, a drying or wetting signal was evident according to the signal-to-noise ratio for some ensembles but not for others. For example, the QldFCP-2 ensemble shows widespread agreement for decreased mean and p99 along coastal regions of north-eastern Australia, while the BARPA-ACS ensemble shows widespread agreement for increases over northern Australia, neither of which is reflected in the outputs for the other modelling groups. Similarly, over southeast Australia, all model ensembles except for NARCliM2.0 show widespread agreement for increased p99.7. These results indicate that relying on a single ensemble of downscaled projections may give a false sense of the certainty of precipitation changes.

Fig. 3: Comparison of ensemble mean change to annual daily mean, 99th percentile, and 99.7th percentile precipitation per degree of global temperature change derived from SSP370.
figure 3

Highlighted values show the average change over Australia in percent, stippling shows where the signal-to-noise ratio >1.0, and boxplots show the variability in the average change over Australia from each of the model ensembles.

Full size image

All model ensembles show precipitation to vary according to the intensity of the precipitation metric assessed. The smallest (greatest) increases (decreases) were seen for mean precipitation, while the largest increases were noted for the most extreme precipitation events (1 in 100 AEP event). Ensemble mean precipitation was projected to change by between −2.5% K−1 and 2.6% K−1, while p99.7 changed by between −0.1% K−1 and 4% K−1, depending on the ensemble used (Fig. 3). These changes are greater when evaluating the response of rare extremes to climate change (Fig. 4), where ensemble median changes ranged from 2.6% K−1 to 6.5% K−1 for the 1 in 10 AEP event and from 4.5% K−1 to 10.1% K−1 for the 1 in 100 AEP event. Both NARCliM2.0 and QldFCP-2 projected smaller increases in rare precipitation extremes over Australia than the CMIP6 host models (Fig. 4). By contrast, CCAM-ACS and BARPA-ACS projected increases greater than the CMIP6 host models. The increases from most ensembles tended to be greatest over northern Australia, especially from the BARPA-ACS ensemble. Despite this inter-ensemble agreement, there were few areas in any ensemble that showed agreement for the 1 in 100 AEP according to the signal-to-noise ratio, due to the variability of future extremes. Corresponding changes to the spatial patterns of the generalised extreme value (GEV) fitted parameter values are presented in the Supplementary Materials (Fig. S1).

Fig. 4: Comparison of ensemble median change to 1 in 10, 50, and 100 AEP daily precipitation per degree of global temperature change derived from SSP370.
figure 4

Highlighted values show the average changes over Australia in percent from the multi-model median, stippling shows where the signal-to-noise ratio >1.0, and boxplots show the variability of the median change over Australia from each of the model ensembles.

Full size image

Over Australia, the intra-model variability tended to be greater within the NARCliM2.0 ensemble than the other modelling groups, despite sampling the least GCMs (Fig. 3). This is especially evident for p99.7 and appears to be due to the very wet model (EC-Earth3-veg) and very dry (ACCESS-ESM) becoming more wet and dry following downscaling (Fig. 5). By contrast, the other downscaling methods tended to bring these outlier models in towards the mean, with some exceptions, including EC-Earth3 downscaled with BARPA-ACS which showed the largest increases of any model assessed for both the 99.7th percentile and the 1 in 100 AEP event (Fig. 5). The QldFCP-2 ensemble tended to have the least variability in the projections, followed by the CCAM-ACS. As both ensembles share the same model for downscaling, this could relate to the CCAM model effectively dampening the magnitude of changes from outlier models.

Fig. 5: Projected changes to common and rare extreme precipitation events for each of the individual climate models against projected regional warming at the end of the century (2071–2100).
figure 5

Changes to a 99.7th percentile (mean) and the (b) 1 in 100 AEP (median) shown. Models are grouped by ensemble according to colour, with the letter corresponding to the host model. Ensemble averages are shown using + symbols. Boxplots show the variability of all host and downscaled models.

Full size image

There was widespread agreement from all ensembles for a decrease to mean precipitation (−11.2% K−1 to −8.6% K−1), p99 (−5.6% K−1 to −4.2% K−1), and p99.7 (−3.6% K−1 to −1.5% K−1) in Perth (Fig. 6) with smaller decreases for more intense events compared to the mean. In other capital cities, there was less consensus in the magnitude or sign of change, though generally decreases to mean precipitation were projected from most ensembles in Adelaide and Melbourne, with only CCAM-ACS tending towards a slight increase. Most capital cities tended towards an increase in the more intense events (p99.7) compared to the mean, with the largest increases typically noted for Darwin (-2.3% K−1 to 8.3% K−1). In comparison to the projections for the means and common extremes, rare extremes tended towards an increase for almost all ensembles in all greater capital cities. Darwin has the largest ensemble median increases in the 1 in 100 AEP event (8.3% K−1 to 18.2% K−1), with the greatest increases reported from the BARPA ensemble (Fig. 7).

Fig. 6: Changes to daily mean, 99th percentile, and 99.7th percentile precipitation per degree of global temperature change derived from SSP370.
figure 6

Boxplots show variability in projected changes within greater capital city regions from each of the model ensembles.

Full size image
Fig. 7: Changes to 1 in 10, 50, and 100 AEP daily precipitation per degree of global temperature change derived from SSP370.
figure 7

Boxplots show variability in projected changes within greater capital city regions from each of the model ensembles.

Full size image

Here, the BARPA ensemble appears to show greater projected changes to extremes compared to the other CORDEX ensembles (Fig. 7). Similarly, NARCliM2.0 appears to show lower projected changes in Melbourne, while QldFCP-2 shows smaller changes in Brisbane when compared to the other CORDEX ensembles. In general, there were considerable differences in the projections across all capital cities, particularly those in northern and eastern Australia (Brisbane, Darwin, Sydney, ACT, and Melbourne). This variability is evident when comparing the distribution of the combined downscaled CORDEX ensemble to the individual downscaled ensembles (Figs. 6 and 7), highlighting the uncertainty associated with projections of precipitation and the need to adopt large ensembles using different downscaling methodologies to quantify the uncertainty. Generally, the combined downscaled ensemble appears to have a similar spread of projected changes to the host model ensemble, except in those regions discussed above (Figs. 7 and 8).

Fig. 8: Probability density function plot of change to daily 1 in 10, 50, and 100 AEP daily precipitation per degree of global temperature change derived from SSP370.
figure 8

Coloured bars show median change for each of the individual downscaled ensembles, while dotted vertical lines show the host and weighted downscaled model ensemble averages (each modelling group weighted evenly). Bold blue and red text shows projected changes per degree warming from the host and downscaled ensembles, respectively.

Full size image

The median projection of change from the host model ensemble closely matched that from the weighted downscaled ensemble (each modelling group weighted evenly; Fig. 8 and Table S4). This was, however, not true for Adelaide, where the host models showed 1 in 100 AEP was projected to increase by 9% K−1 compared to 4% K−1 from the downscaled ensemble or Brisbane, where the host models showed a 3.7% K−1 compared to 7.6% K−1 from the downscaled ensemble. Both groups projected the largest increases to the 1 in 100 AEP for Darwin (12.2% K−1 and 11.9% K−1 for the host and downscaled ensembles, respectively), which is considerably larger than what would be expected from the CC relationship (Fig. 8).

Discussion

Our study shows that downscaling consistently reduces biases of common extreme precipitation (p99 and p99.7) over capital city regions in Australia when compared to the host CMIP6 GCMs (Fig. 1 and Table S2 to S3). In contrast, the host models performed comparably well for mean precipitation within the capital city regions and for mean and common extreme precipitation when assessed across Australia as a whole. These differences relate to the coarse model resolution of the GCMs, which are not able to represent precipitation patterns and extremes over regions with complex terrain, land use changes, or coastal gradients when compared to finer resolved RCMs18,19,20. It is, however, within these complex regions where the majority of the Australian population resides and where the performance differences between host and downscaled models are greatest. RCMs, therefore, appear to be better suited to provide information on climate hazards within these populated areas, particularly for extremes.

There were considerable differences in the spatial extent of biases from the different downscaled modelling ensembles and individual models. Both NARCliM2.0 and QldFCP-2 had notable dry biases over Australia and the capital city regions, while CCAM-ACS and BARPA-ACS had notable wet biases, though to a lesser extent. The dry biases from the QldFCP-2 ensemble are particularly evident over northern Australia and may relate to a misrepresentation of the number and intensity of low-pressure systems and cyclones, which are major contributors of precipitation, especially extremes in these regions37. NARCliM2.0 shows similar dry biases over northern Australia and within the greater capital city regions (Fig. 2) for mean and p99, but much less bias for p99.7. Di Virgilio et al.34 posited that the dry bias over northern Australia could relate to issues in capturing the Australian monsoon. They also showed that there was less bias in mean precipitation over southeast Australia from NARCliM2.0 compared to previous downscaled models from CMIP5 (NARCliM1.5) and CMIP3 (NARCliM1.0). By contrast, the wet bias present in the CCAM-ACS ensemble has been shown to relate to an overestimation of extreme precipitation, while low precipitation events were underestimated from this ensemble36. These biases are believed to relate to the parameterisation of CCAM, with improved schemes currently being tested to resolve these issues. Similarly, the wet bias from the BARPA ensemble has been shown to relate to a general overestimation of precipitation extremes and the inclusion of too many low-intensity wet days. It is interesting to note the different signs of the biases between the CCAM-ACS and the QldFCP-2 as both ensembles make use of the CCAM model for downscaling. These differences likely relate to the downscaling approach adopted, particularly in regard to sea surface temperatures (SSTs), which have a significant influence on precipitation. Here, QldFCP-2 used bias-corrected sea-surface temperatures from the host models as a main forcing, whereas CCAM-ACS adopted a nudging approach for the atmosphere and uncorrected host model sea-surface temperatures, leading to diverging simulations of precipitation from the same host model.

Annual mean precipitation has been noted to be increasing over parts of northern Australia, and decreasing over southern parts, with the greatest decreases noted along the southeast and southwest coasts, where a decline approaching 10% per decade has been observed38. Over other regions, the trends in mean precipitation are less clear. Comparisons of the long-term trends have shown downscaled models to generally follow the host models but compare well against observations where the trends were significant39. Interestingly, in regions with significant trends in the observations, projections of future mean precipitation can be seen to lead to further wetting or drying from most ensembles (Fig. 3). Typically, extreme precipitation occurs in summer in northern Australia, and in winter/spring in southern Australia, although the timing is highly variable in southern Australia with extremes occurring throughout the year40. Seasonal precipitation has been shown to be well represented by the different downscaled ensembles considered in this study20,34,39,41, with greater relative biases generally noted for winter compared to summer39. Using the QldFCP-2, Chapman et al.20 showed that downscaling led to improved representation of the seasonality of precipitation for a majority of the models, with an ensemble average improvement of 13% noted. Considering all four ensembles, Jiang et al.39 showed benchmarks of seasonality to have been met for all ensembles in most regions, with issues mainly noted for southern Australia.

Extreme precipitation was projected to increase across Australia and all capital cities, with greater increases seen for rarer extremes compared to more common extremes (Fig. 9), in line with previous studies13,14,15,16. The magnitude of the projected changes was, however, dependent on the combination of the model ensemble and region considered. We found that changes for the 99.7th percentile precipitation ranged from between −0.1% K−1 to 4.0% K−1 (Fig. 3), while changes to the 1 in 100 AEP ranged from between 4.5% K−1 to 10.1% K−1 across Australia (Fig. 4), which are the result of increases to all GEV parameters (Figs. S1 and S2). However, there was generally high uncertainty associated with these projections, with few regions showing significant changes following our signal-to-noise ratio analysis. These projected changes are broadly in line with the findings of Wasko, Westra, et al.42, who suggested a scaling rate of 8% K−1 for Australia from a meta-analysis of available studies of observations and projections. Observations have pointed to precipitation becoming more episodic throughout Australia43, with an increase in precipitation intensities noted, particularly for rarer sub-daily events44,45,46. To date, available studies on projections have made it difficult to ascertain if there are any geographic differences in these scaling rates42, as the majority of the literature to date has focused on the populated southeast Australia29,30, which have similarly pointed towards scaling rates exceeding the CC relationship (>7% K−1).

Fig. 9: Summary of projected changes to mean and extreme precipitation from downscaled and host models for greater capital city regions.
figure 9

Background shows the mean precipitation change per degree global warming across Australia from the full ensemble of all downscaled projections (each modelling group weighted evenly). Changes by greater capital city region for: a Darwin, b Brisbane, c Sydney, d Australian Capital Territory, e Melbourne, f Hobart, g Adelaide, and h Perth. Error bars show the 10th and 90th percentile of changes.

Full size image

Using the largest ensemble of projections to date, we show that there do appear to be large-scale geographic differences in projected precipitation extremes, with greater increases generally projected for northern Australia compared to southern Australia. Similar findings have been noted when assessing the observational records2,47 and from a recent study based on 4 downscaled CMIP5 GCMs42. While this north-south difference is generally evident across the ensembles, it is important to note that there are widespread regional variations from the different ensembles for all precipitation intensities, except for southwest Australia where a strong signal of decreasing mean precipitation is shown for all ensembles (Fig. 3). Decreased mean precipitation, however, did not necessarily translate to a decrease in common or rare extremes, with generally consistent increases to rare extremes still projected for Perth despite the declines to mean precipitation (Fig. 9).

In other regions, a drying or wetting signal is evident according to the signal-to-noise ratio for some ensembles but not for others. For example, the QldFCP-2 ensemble reported significant decreases in mean and 99th percentile precipitation for coastal regions of north-eastern Australia, whereas BARPA-ACS noted significant increases over northern Australia, neither of which is reflected in other modelling group outputs. Relying on a single ensemble of projections from a single modelling group could therefore give a false sense of the certainty of the climate change signal for extreme precipitation events. These results are consistent with previous studies, which have shown downscaling methodologies to be a significant source of uncertainty to multiple climate metrics, including precipitation extremes48,49. There is a clear need to consider multiple ensembles of projections derived from multiple downscaling methodologies to account for this uncertainty, particularly if modelled outputs are to be used by decision makers.

Additional sources of uncertainty include natural variability, scenario uncertainty, and epistemic uncertainty, which have not been explicitly considered in this study. Natural variability has been shown to be the largest source of uncertainty for projections of mean and common precipitation extremes globally48, including over Australia, which is subject to very high inter-annual variability. While the consideration of a large ensemble of projections goes some way to accounting for natural variability, quantifying this uncertainty would require multiple ensemble members from individual models. Scenario uncertainty results from the plausible range of future possible emissions pathways and has been shown to be a comparatively minor source of uncertainty for precipitation extremes48,50. The estimates of changes per degree of global warming used in this study have all been derived from the SSP370 emissions scenario, as this was the highest emissions scenario shared by all the ensembles. Aerosols and land use changes from this scenario differ considerably from the other scenarios used for impact assessments (SSP126, 245, and 585), which may contribute to an underestimation of precipitation for some regions, particularly Asia51. However, while forcing scenarios have been shown to be influential for scaling rates of mean precipitation52,53, their influence has been shown to be indistinguishable for precipitation extremes which are instead much more reliant on atmospheric warming rates16,52,53,54. Nonetheless, future work could compare the mean and common extreme precipitation from this study to other emissions scenarios (e.g., SSP585) to determine if these differences are significant. Lastly, there is a degree of epistemic uncertainty associated with the estimates of extremes, particularly for the rarest extremes (1-in−100 AEP)55. These are the result of fitting distributions to a relatively limited dataset (30 years), though pooling may go some way to helping reduce this16. Additional epistemic uncertainty in projected precipitation changes is present due to incomplete knowledge on how to represent precipitation processes (parameterisation of clouds, convection and turbulence) in climate models under a warming environment.

Projected changes to extreme precipitation are the result of thermodynamic and dynamic processes. Thermodynamic processes have been shown to lead to an increase in precipitation extremes in the order of 4 to 8% per degree warming globally10, while dynamic processes, which influence the frequency and intensity of synoptic and subsynoptic features, can increase or decrease this change but have high spatial variability and uncertainty. Robust reductions in the contribution of dynamic processes to extremes have been projected for subtropical regions, including parts of Australia10, which may explain the reduced rate of increase in moderately extreme precipitation shown in this study, when compared to other regions16. The rarer events consistently have higher per-degree changes, suggesting that they have larger contributions from dynamic processes. That is, the synoptic situation needs to be acted upon to enhance the thermodynamic effect in order to generate these rare extremes.

In northern Australia, tropical cyclones are an important contributor of extreme precipitation, especially in the northwest, where over 40% of extreme rainfall days (above the 99th percentile) are estimated to coincide with a tropical cyclone day37. Projections have pointed towards a reduction in the number of tropical cyclones impacting Australia, particularly the northwest, with less certain changes for the north central and northeast56,57. Reductions to the number of cyclones impacting northern Australia may help explain reductions to the moderate extremes noted for parts of northern Australia from some ensembles. However, some studies have suggested that there will be an increase in the intensity of these events, possibly contributing to increased precipitation extremes58. Extratropical cyclones are a major contributing factor for precipitation extremes along Eastern Australia59,60 and will likely become less frequent in the future, contributing to mean precipitation decreases. However, precipitation extremes associated with these events are shown to increase roughly in line with CC relationship60,61.

Convective thunderstorms are important contributors of extreme precipitation across Australia62, they have been observed to increase in intensity over recent decades, and are expected to intensify further due to climate change63. However, the dynamically downscaled models used in this study are resolved at spatial scales between 10 and 20 km, which are not adequate to explicitly represent convective processes. Development of very high resolution models (<4 km), which are able to explicitly represent convection is currently ongoing at a regional scale34,64, and may lead to improved simulation of precipitation extremes, particularly for rare extremes65,66,67,68. Benefits of convective permitting models would be especially notable for sub-daily precipitation extremes, which are more closely linked to convective processes such as thunderstorms. These sub-daily extremes have been shown to be increasing at a rate exceeding daily extremes4,42,45,69, which will have greater ramifications for flooding and infrastructure impacted by short-duration rainfalls such as stormwater systems.

Our results show increases in extreme daily precipitation events are likely across Australia and its greater capital city regions, elevating the risk of flood events. Across Australia, flooding is already the costliest natural disaster70, impacting all regions and population centres71,72. Within cities, these increases may be compounded by continual urban expansion, which increases runoff and exacerbates flooding73. Urban expansion and population growth also work to increase the exposure risk to flooding74,75. This trend is evident in the recent past, with a near doubling of the global urban area in floodplains impacted by 1 in 100 AEP events between 1985 and 201576. Continual population growth and urban expansion may therefore work synergistically with climate change to exacerbate not only the magnitude of flooding events but also the population at risk and the cost of the potential damages. These global issues necessitate concerted adaptation and planning measures to mitigate future development within at-risk floodplains and to improve resilience.

In some cases, increases to the magnitude of the largest flooding events may see a push for upstream dams to be increasingly used for flood mitigation instead of water supply77, which would necessitate dams to operate at lower maximum storage to accommodate larger flow volumes. Projected mean rainfall declines and the subsequent declines in the smaller more frequent flood events, which are important for water supply78 could also reduce water security for some regions. Dam managers may find it increasingly hard to prioritise flood mitigation at the same time as water security, as these two priorities become increasingly at odds with one another. Concurrently, more intense precipitation will likely exacerbate water quality issues by elevating erosion and nutrient runoff from agricultural lands79, which can have ecological impacts and cause drinking water supply issues. The projected changes to precipitation extremes highlighted in this study will therefore have a host of ramifications for flooding, the environment, dam management, water supply, and agriculture within Australia.

To conclude, our analysis explores a large ensemble of CORDEX-CMIP6 regional projections, including multiple host CMIP6 models to understand changes in regional extreme precipitation. It utilises an innovative approach combining dynamical downscaling, generalised extreme value distribution and global warming level analysis to unravel the impacts of climate change in rare extreme precipitation events across greater capital cities, where approximately two-thirds of the Australian population reside. The findings revealed globally relevant scientific insights and can inform decision-making around topical issues such as flood management, urban water supply, urban planning, and agriculture.

Methods

Study area

This study evaluated changes to extreme precipitation across the Australian continent, which encompasses a range of climate regions, including arid, equatorial, savannah, subtropical, temperate, and tropical regions. We further examined the changes within greater capital city regions, where approximately two-thirds of the Australian population lives (Fig. 10).

Fig. 10: Location of the study extent and a comparison of how the GCMs and RCMs consider topography compared to actual topographic values.
figure 10

Extent of study area and location of Australian greater capital city areas adopted in this study compared to major climate regions (a). Comparison of the resolution of Digital Elevation Models (DEMs) over Sydney from the Shuttle Radar Topography Mission (SRTM) (b), the QldFCP-2 RCM (c), and the ACCESS_CM2 GCM (d). Cross-sectional comparison of these DEMs across the Great Dividing Range (e), highlighting improvements in the RCM over the GCM.

Full size image

Data

We used a combined ensemble of high-resolution dynamically downscaled climate simulations for Australia generated by four Australian modelling groups. The four modelling groups applied three independent RCMs for dynamical downscaling, used in five different configurations. The GCMs selected for downscaling were based on different selection criteria80,81,82, which aimed to best represent the future spread in the climate change signal from the ensemble of CMIP6 models, while prioritising models which were statistically independent and better able to represent the Australian climate. In total, 19 different CMIP6 GCMs were chosen for downscaling, with some GCMs downscaled multiple times by different RCM configurations (Table 1).

QldFCP-2 and CCAM-ACS used the global stretched grid Conformal Cubic Atmospheric Model CCAM; Thatcher, 2020) for downscaling 11 and 7 GCMs, respectively. QldFCP-2 downscaled to a 10 km spatial resolution over Australia, and used bias and variance corrected SSTs and sea ice, following the approach outlined by Hoffman et al.83. Five of the QldFCP-2 CCAM simulations were run using dynamic atmosphere-ocean coupling, while the rest were run in atmosphere-only mode (Table 1). CCAM-ACS downscaled to a 12.5 km spatial resolution and employed spectral nudging to constrain the model to follow the host GCM84, with all models run using dynamic atmosphere-ocean coupling. Both QldFPC-2 and CCAM-ACS used similar but not identical configurations for atmosphere, ocean, land-surface and aerosol parametrizations, and so the differences between them are mainly due to downscaling design (spectral nudging vs bias-corrected SSTs and sea ice), different GCMs, and differences in resolution. The Bureau of Meteorology Atmospheric Regional Projections for Australia (BARPA) is an RCM based on the UK Met Office Unified Model and Joint UK Land Environment Simulator (JULES) but configured for Australia and has been applied to downscale 7 GCMs to a 17 km spatial resolution85. Similar to CCAM-ACS, nudging was used to constrain the model to follow the host GCM86. NARCliM2.0 (New South Wales and Australian Regional Climate Modelling) employed two configurations of the Weather Research and Forecasting (WRF) model, each adopting different parameterisations of selected physics options to downscale 5 GCMs. The two different sets of RCM parameterisations were selected based on their ability to simulate Australia’s recent climate and statistical independence from a larger set of 78 structurally different configurations. NARCliM2.0 provides simulations at a 20 km spatial resolution over Australasia. BARPA and NARCliM2.0 follow a limited area modelling approach, and as such, these RCMs were forced at the lateral boundaries and used SST from the host models. In total, a large ensemble consisting of 39 different regional models was considered in this study across the 4 modelling groups. Further details on the individual downscaling experiment designs can be found in20,34,35,36.

Daily observed gridded precipitation data with a spatial resolution of 0.05° (~5 km) were obtained from the Australian Gridded Climate Data Project (AGCD87). Prior to analysis, all datasets, including GCMs, RCMs, and observations, were re-gridded to the same spatial resolution (i.e., 10 km) using distance weighting interpolation. We assessed the performance of host and downscaled models against the observational data using daily mean, 99th percentile precipitation (p99), and 99.7th percentile precipitation (p99.7). Regions with poor observational data quality during the comparison period were masked out of the model evaluation (Fig. 1). Model performance using these metrics was evaluated over Australia and over the eight greater capital city regions (Fig. 10) using the MAPE and the RMSE to quantify differences.

Extreme value analysis

Extreme value analysis was applied to assess changes to the probability distribution of rare extreme events. Here, these events represent the 1 in 10, 1 in 50, and 1 in 100 AEP, which approximately correspond to events with annual return intervals of 10 years, 50 years, and 100 years, respectively. We sampled the daily timeseries of precipitation at each grid cell using the block maxima approach to derive AM precipitation and then pooled together data from nearby cells using a 5 × 5 box centred on each grid cell to extend the data series used for the extreme event analysis88. For the analysis of capital cities, we chose to pool together all the AM data within each of the regions for the analysis. It is assumed that all these regions are homogeneous for the purpose of pooling. Due to weak spatial dependence of precipitation extremes, similar pooling approaches adopted in the literature have been shown to help reduce the uncertainty of local estimates of the GEV parameters16,89. For instance, the index flood approach90 has long been usefully applied for flood frequency analysis. The GEV distribution was then fitted to the AM series using the L-moments method for parameter estimation, which is shown to be more robust than the maximum likelihood estimation approach for relatively short samples91. The GEV distribution is a generalised expression combining the Gumbel, Fréchet, and the Weibull distributions and is given by:

$$G\left(x\right)=\exp \left\{-{\left[1+\xi \left(\frac{x-\mu }{\sigma }\right)\right]}^{\frac{-1}{\xi }}\right\},{for}\left\{x:1+\xi \left(\frac{x-\mu }{\sigma }\right) > 0\right\}$$
(1)

Here, μ, σ, and ξ are the location, scale, and shape parameters, respectively. Here, the location parameter is a measure of the central tendency and is loosely linked to the mean, the scale parameter is a measure of variance, and the shape parameter describes the tail behaviour.

The shape parameter is particularly important for the estimation of rare extremes, as it describes the behaviour of the tail. When ξ ≥ 0, the distribution is unbounded with no upper limit, while when ξ < 0, the distribution is bounded by an upper limit92. Accurate estimation of the shape parameter necessitates many years of data to adequately fit, as it is susceptible to outliers93. We fit the GEV distribution to two 30-year periods representing the recent past (1981–2010), which we term the reference period, and the far future (2071–2100). However, as the data are pooled from nearby cells, we effectively increase the number of data points used in the analysis from 30 per grid cell to 75088. In accordance with extreme value theory, it is assumed that the data remain relatively stationary within the 30-year periods. This is a commonly held necessary assumption, as nonstationary models cannot be reliably fit to short periods16.

Climate change assessment

We examined the impacts of climate change on precipitation extremes by the end of the century (2071–2100) relative to the 1981–2010 reference period. We calculated the precipitation for the 1 in 10, 50, and 100 AEPs. We also evaluated changes to mean and common extreme (p99 and p99.7) precipitation to determine changes across a range of precipitation intensities. The 99.7th percentile precipitation was evaluated as this approximately corresponds to an event that would occur once per year. Calculations were applied at each individual grid cell for each of the climate models considered. The results for each of the projections were assessed individually and by model ensemble (i.e., four groups of RCMs and one group of GCMs), including a combined ensemble of all downscaled projections (CORDEX ensemble) assessed for the greater capital city regions. Model ensemble averages and medians were calculated, with calculations of the averages adopting a one-model-one-vote rule, resulting in an 11-model average for QldFCP-2, a 5-model average for NARCliM2.0, a 14-model average for the CMIP6 GCMs, and a 7-model average for CCAM-ACS and BARPA-ACS (Table 1). Resulting changes are presented as spatial maps over Australia using the ensemble average or median change and as boxplots for the eight greater capital city regions assessed using all model projections to better understand uncertainty. Averages were used for the analysis of mean, p99, and p99.7, while medians were used for the results of the GEV extremes to ensure they were not influenced by outliers.

We present precipitation changes as a rate per degree of global temperature change (i.e., the global mean temperature, including both land and ocean regions). For consistency, we used the global temperature changes derived from the host GCMs as opposed to warming from the downscaled projections. This was implemented to ensure consistency between the different model ensembles and due to the difficulty in deriving global warming levels from limited area RCMs such as BARPA and WRF. We calculated the projected changes per degree warming for the high emissions scenario (SSP370) only, as only this scenario and SSP126 were shared by all modelling groups, and as SSP370 would show a more sensitive response by the end of the century. Scaling rates were calculated for each model from the projected changes between the reference period (1981–2010) and the end of the century (2071–2100) as per the equation below.

$$\frac{ \% {change}}{{degree\; warming}}=\left[\left\{{\left(\frac{{P}_{f}}{{P}_{r}}\right)}^{\left(\frac{1}{{T}_{f}-{T}_{r}}\right)}\right\}-1\right]\times 100\,$$
(2)

Where, Pf is the precipitation in the future period, Pr is the precipitation in the reference period, Tf is the mean global temperature in the future period, and Tr is the mean global temperature in the reference period. To determine where there is confidence in the scaling rate change over Australia, we adopt the signal-to-noise ratio to see where the climate change signal emerges over the ‘noise’ for each of the climate model ensembles considered94. Here, we consider the noise as the standard deviation of all the models in each ensemble. Stippling is shown on the ensemble mean and median change maps where the signal-to-noise ratio is >1.0, as this is a commonly adopted threshold used in the literature94,95. To compare differences between the downscaled models and host models, we also compared the probability density function plot of the changes from these RCMs and GCMs.