Sea surface temperature driven modulation of decadal co-variability in mean and extreme precipitation

This study investigates the role that sea surface temperature (SST) variability plays in modulating the relationship between decadal-scale mean precipitation and monthly-scale extreme precipitation using the Australian Community Climate and Earth System Simulator Earth System model (ACCESS ESM1.5) climate model. The model large ensemble successfully reproduces the observed strong co-variability between monthly mean rainfall and wet extreme rainfall, defined as monthly rainfall totals above the 95th percentile. Removing SST variability in the ACCESS ESM1.5 model significantly weakens the co-variability between mean and wet extremes over most of the globe, showing that SSTs play a key role in modulating this co-variability. The study identifies Pacific and Atlantic SST patterns as the main drivers of the decadal scale co-variability in mean and extreme wet precipitation. On the other hand, observations and model results show that co-variability between mean and dry extremes is generally weaker than for wet extremes, with highly regional signals. Model experiments also show that SST variability plays a weaker role in modulating the co-variability between the mean precipitation and dry extremes as compared to wet extremes. These results suggest that stochastic atmospheric variability plays a stronger role in generating dry precipitation extremes compared SST forcing.


Introduction
Many significant global climate-related impacts and associated economic losses have been associated with precipitation variability (Dike et al 2022a).High precipitation events lead to an increased incidence of flood disasters that affect local populations and critical infrastructure (Zhang et al 2018, Dike et al 2022b).Conversely, high-severity droughts (low precipitation events) threaten agriculture (Sun et al 2012, FAO 2015, Ukkola et al 2020, Ayugi et al 2022, Navarro and Merino 2022).Notably, both observational and numerical model studies indicate that precipitation extremes, like the above, do not necessarily occur in isolation, and there is clear spatio-temporal variability on the inter-annual time scale (Thompson and Wallace 2001, Zhang and Chan 2009, Smith and Johnson 2010, Williams et al 2012, Wang and Lee 2013, Davis et al 2016, Johnson et al 2017, Garcia and Rodriguez 2018).
The frequency and intensity of precipitation extremes can manifest across different time scales.Variations can occur in the short-term (e.g.monthly) and long-term (e.g.decadal) (Taschetto et al 2016, Seager and Ting 2017, Adamu et al 2022, Dike et al 2022b).Large precipitation deficits on monthly time scales lead to drought on this time scale, affecting sectors such as agriculture (Meroni et al 2017, Singh et al 2019, Mishra and Shah 2019, Miao et al 2016).Significant precipitation deficits also occur on decadal-scale time scales but are smaller in magnitude because of a longer averaging period.However, the long duration of these average deficits means that they significantly impact water resources (Power et al 1997, Verdon-Kidd and Kiem 2009, Cook et al 2014, Henley et al 2015, Seager and Ting 2017, King et al 2020).
Decadal and monthly-scale covariability in precipitation is particularly important as it pertains to precipitation extremes.For example, Adamu et al (2022) showed that decadal-scale low mean precipitation was largely unrelated to monthly-scale drought variability.Instead, the study showed that this low decadal-scale mean rainfall was more closely linked to variations in monthly-scale wet extremes.Moreover, the study showed that at many global locations, decadal mean precipitation was largely modulated by variability in wet monthly-scale preciptiation extremes.
Previous research suggests that variations in sea surface temperatures (SSTs) can influence both wet and dry extremes at monthly to interannual time scales over various regions of the globe.SST variability stemming from the Pacific and Atlantic regions was identified as particularly important in modulating either mean or extremes (McCabe et al 2008, Mohino et al 2011, Nnamchi et al 2013, Lin and Dike 2018, 2018, Feng et al 2022).Research by Feng et al (2022) revealed that decadal-scale precipitation variability is influenced by shifts in entire distribution patterns.McCabe et al (2008) further underscored the substantial impact of decadal-scale precipitation variability, elucidated shifts in entire distribution patterns and emphasized the role of SSTs in this dynamic process.While this research is important, the majority of studies either focus on mean (i.e., decadal-scale) or extreme (i.e.monthly) precipitation in isolation leading to a gap in how SST variability influences covariability of changes in mean and extremes.
Although the connections between SST forcing and to decadal-scale extremes are well documented, there is still debate as to the magnitude of the role of SSTs.Cook et al (2011) andSeager andTing (2017) have shown a role for a combination of SST patterns in the Pacific and Atlantic oceans, combined with stochastic variability, driving droughts in the southwest USA.They also argued that the regions' long-term droughts are primarily SST forced.Despite showing decadal-scale Australian droughts are possible in the absence of oceanic variability, Taschetto et al (2016) also argued for the strong role of SST variability by highlighting that oceanic variability regulates the magnitude and frequency of long-term wet and dry spells in Australia.However, Stevenson et al (2015) used modeling experiments to describe show that megadrought in North America is chiefly controlled by stochastic variability and land-surface feedbacks, not SST variability.Thus, the role that SST variability plays in regulating decadal-scale extremes, and how this is modulated through changes to monthly-scale extremes, is still the subject of some debate.
Motivated by these gaps in understanding, this paper aims to systematically explore how SST variability modulates the intricate relationship between decadal-scale mean precipitation and with monthly-scale dry and wet precipitation extremes.Leveraging an ensemble of Australian Community Climate and Earth System Simulator Earth System model (ACCESS ESM1.5) historical simulations and an idealized experiment, this research endeavors to contribute to a more comprehensive understanding of the complex interplay between SSTs variability, mean precipitation, and extreme events, thereby advancing our capacity to anticipate and adapt to changing climate patterns.

Historical simulations
The ACCESS ESM1.5 large ensemble of the CMIP6 historical simulation, developed for community use by the ACCESS development team, contains 40 members.The historical simulations, which extend from 1850-2014, are forced by observed estimates of radiative forcing (i.e.GHG concentrations, solar forcing, land use changes, etc.) as described by Eyring et al (2016).
Here, we utilise monthly mean precipitation (pr) and surface temperature (tas) from all 40 simulations spanning 1870-2013.All outputs were processed on their native grid.We then assess the ability of the The simulation was run through the same period as the historical simulations (1870-2013), and the same radiative forcing was applied.As this is an AGCM simulation, SSTs were applied as a model forcing calculated from the monthly SST data of the 40 historical simulation members spanning 1870-2017.
In order to remove SST variability from these forcing SSTs, the forcing SSTs were generated by taking the average of the monthly SST values across all 40 historical members.As the internal variability of SST in each ensemble member is considered independent, the variance for an ensemble average SST is shown to scale by 1/N, where N is the number of ensemble members (Liguori et al 2020).As such, the ensemble mean SSTs would contain a forced signal in response to applied radiative forcing changes (i.e. a seasonal cycle and anthropogenic warming), while the internal variability is approximately 1/40th the magnitude of an individual ensemble member.The coupling between the land component and atmospheric fluxes in this experiment remains active.By removing SST variability and allowing other components of the atmosphere to interact, this experiment aims to investigate whether SST variability modulates the relationship between the mean precipitation on decadal time scales and wet extremes, later defined, on monthly time scales.

Atmosphere only SSTA forced experiment (AGCM SSTA )
A second ACCESS ESM1.5 atmosphere-only experiment was conducted to bolster confidence in comparing the results between the historical ensemble and our AGCM SSTA removed experiment.The experiment is forced with SSTs and sea ice from a randomly selected member of the 40-member historical ensemble, hereafter AGCM SSTA Thus, this AGCM experiment includes the SST of one historical ensemble member (i.e., restoring SST variability in the atmosphere-only simulation).The idea of using one ensemble member or single observation data for assessing climate model simulations is supported by several studies.For example, Hausfather et al (2017), found that, in certain situations, a single simulation adequately captures observed climatological statistics.Other studies, like those by IPCC ( 2013) and Eyring et al (2016), have taken a similar route, focusing on evaluating climate models using observations.For our experiment, we are expecting the results to align with what the large ensemble suggests.

Methodology
We first used ACCESS ESM1.5 ensemble data to evaluate whether the coupled climate model can reproduce observed relationships between precipitation means and extremes.Wet and dry extremes of monthly precipitation are defined as 95th (P95th) and 5th (P5th), respectively, while decadal variability is represented by computing these values along with the mean in a 10 year running window.Prior to this calculation, data are split into cool and warm seasons of April-September and October-March for the southern hemisphere, which is inverted for the northern hemisphere.Decadal mean precipitation (Pmean) was serially correlated with decadal wet and dry extremes at each grid point to assess spatial covariability between mean and extremes.
To further investigate the structure of SSTs that influence the relationship between decadal mean precipitation and wet extremes, we employ principal component analysis (PCA) to identify prominent spatial patterns of precipitation mean and extreme co-variability (Bretherton et al 1992, McGregor andHolbrook 2008).Here, we focus on mean and 95th percentile only as these show the most coherent covariability, as guided by analysis from Adamu et al (2022).
In computing the PCA, we subtract the ensemble mean (i.e., forced signal) from each ensemble member to focus the PCA on the internal variability of the of 95th percentiles and the mean of precipitation.As each data point (or spatial map) represents precipitation in a 10 year sliding window, we skip 18time steps (3 years) between data points to ensure that 30% of data in each sliding window calculation is independent of surrounding data.We then concatenate all historical ensemble members to form a large dataset with dimensions (space, time, 40 ensemble members), such that the PCA produces one set of spatial patterns for the entire ensemble.From PCA, the higher order of outputted pairs of spatial maps (one for the mean and the other for 95%), more mean-squared temporal co-variability is explained.Each set of maps also has a pair of PCA expansion coefficients, which are highly correlated with each and here referred to as the principal component (PC) time series, that can be used to explain temporal variability of related pairs of patterns.
The PC time series for the mean and 95th percentile is then regressed against decadal mean SSTs to identify if a consistent SST pattern is associated with this co-variability.As with precipitation data prior to this regression: (i) decadal mean SSTs are calculated a 10 yr sliding window, and 18-time steps (3 years) are skipped to ensure each map is generated with at least 30% new data; and (ii) data from all 40 ensemble members is concatenated.

Representation in ACCESS ESM1.5
Ensemble mean correlation maps computed between decadal running means, 95th percentiles (wet extremes), and 5th percentiles (dry extremes) for 40 historical ensemble members are show in figure 1 and table 1.Strong relationships are evident between mean and wet extremes, relatively weak relationships are evident between mean and dry extremes.There is almost no relationship between wet and dry extremes.Averaged over the globe, decadal mean precipitation explains 13% and 14% of variance in decadal dry extremes during April-September and October-March seasons, respectively.Globally averaged, decadal mean precipitation for wet extremes explains 55% and 54.5% of its variance during April-September and October-March seasons, respectively.There are no significant correlations between wet and dry extremes for most parts of globe, with average explained variance of approximately 0.5% in both seasons.This highlights that decadal variability in monthly precipitation distributions in ACCESS ESM1.5 is manifested as changes to skewness rather than shifts to mean, consistent with observational studies in Adamu et al (2022).
Boxplots of spatial cross-correlations between all individual ensemble members are presented in figure 2 to demonstrate consistency of relationships between means and extremes presented in figure 1.This plot shows consistently strong spatial correlations between decadal mean and wet extremes across the 40-member ensemble, with a median spatial correlation coefficient of about 0.85 during both seasons.The consistency of maps displaying a relationship between decadal mean and dry extremes is weaker, with average values of 0.39 during April-September and 0.56 during October-March (figure 2).There is low spatial consistency across ensemble members for maps of the relationship between wet and dry extremes.These results highlight that individual simulations consistently demonstrate a strong relationship between mean and wet extremes.Hence, this provides confidence that ACCESS ESM1.5 simulations can be used to understand further causes of decadal variability in mean and extremes of monthly precipitation.

Exploring the role of internal SST variability
The role of internal SST variability as a driver of decadal-scale co-variability between means and extremes of monthly precipitation, as described in Adamu et al (2022), is now investigated by comparing results from AGCM SSTA removed experiment with the historical ensemble.As discussed in section 2.1, the AGCM SSTA removed experiment effectively removes internally generated variations of SSTs.As with the historical ensemble, decadal co-variability in monthly precipitation distributions of this experiment are calculated by correlating decadal means and extremes at each grid point (figure 3).
With SST variability removed, figure 3 shows that the decadal-scale co-variability between precipitation means and extremes from figure 1 almost disappears.The strong relationship between mean and wet extremes from the historical simulations weakens to be insignificant over most global land areas when SST variability is removed.The median correlation coefficient for the relationship between mean and wet extremes in figure 3 is less than 0.3, which is much lower than value of >0.6 (figures 1(c) and (d)) from the historical ensemble.
The range of correlations produced from the historical simulations acts as a confidence interval to establish the significance of the difference between AGCM SSTA removed and historical simulations.Table 2 shows that only 24% and 22% of the AGCM SSTA removed global land areas for both April-September and October-March periods, respectively, fall within the confidence interval of the historical simulations.Thus, we can conclude that the differences in co-variability between precipitation means and wet extremes are statistically significantly different between the two experiments over the remaining ∼75% of land regions.Supporting the hypothesis that decadal-scale co-variability between the mean and wet extremes is largely driven by SST variability.
Conversely, the relationship between the decadal mean rainfall and dry extremes, which is relatively weak and noisy in the historical simulations (figures 1(a) and (b)) shows similar results in the AGCM SSTA removed experiment.The similarity between the two experiments shows that the relationship between decadal mean rainfall and dry extremes at most global locations is not driven by SST variability.A lack of co-variability with SST forcing implies that the relationship between decadal mean and dry extremes is more strongly modulated by other processes, for example, stochastic atmospheric processes, for large proportions of the globe.This is supported by findings from the work of Stevenson et al (2015) who showed relationship between decadal mean and dry extremes is regulated by stochastic atmospheric processes.

Experiment verification
Next, we test whether or not the results of section 3.2 are due to the different model setup used for the experiments (e.g., coupled model versus atmosphere only experiments).An AMIP experiment was performed (AGCM SSTA ) that incorporated the SSTs and sea ice forcing from a randomly selected historical ensemble member.With SST variability restored, the modeled relationships between the mean and extremes seen in the historical ensemble (figure 1) are also restored (figure 4).This confirms that the difference between AGCM SSTA removed and the historical ensemble is due to the different SSTs used to force the experiments rather than the experiment model configuration.Table 3 shows that more than 80% of global land area falls within the ensemble range of the historical simulations.Thus, the AGCM SSTA experiment is not statistically significantly different to the historical simulations, for the majority of global land areas.This again bolsters confidence in the result that SST variability is a key driver of the decadal-scale co-variability in mean and wet extreme precipitation worldwide.

Exploring the internal SST variability structure
In sections 3.1-3.3,we identified a role for internal SST variability in modulating the co-variability between decadal-scale mean and wet extreme precipitation.We now examine if there are any coherent SST structures that underpin these results.To do this, we conduct a PCA of the mean and 95th percentile of precipitation and identify the leading mode of variability.We then regress the resulting PCs of these modes onto the SSTs to identify if there are any coherent spatial structures (see sup material).
Figure 5 shows the first PCA spatial pattern of co-variability between the running 10 year mean    It is noted that the modelled precipitation response in some regions (e.g.Australia and China) is inconsistent with the literature from the observations.For instance, the observations typically demonstrate an out-of-phase relationship with the PDO/IPO in Australia and China positive IPO leads to decreases in mean precipitation (Verdon et al 2004b, Dong andDai 2015), rather than the in-phase relationship shown in figure 5. Therefore, we verified our results by computing the IPO Tripole Index from Henley et al (2015) and regressing it against SSTs.The results of this regression (not shown) in the Pacific were similar to that shown in figure 5(e), suggesting that the model does not appear to reproduce the observed precipitation response to the IPO/PDO in these regions.To explore this further in the future, we will seek to understand if the observed relationship between the IPO/PDO and rainfall in Australia and China is possible within individual ensemble members.
Figures 5(b) and (d) display the PCA1 spatial pattern for the October-March season, which explains around 37% of the co-variability between the running 10 year mean and 95th percentile of precipitation.Typically, there are similar mean and extreme precipitation patterns for PCA1, and these changes appear strongest in the Tropics and the Southern Hemisphere (SH).We can see clear precipitation decreases over Australia, western Africa south of the equator, eastern/central South America (largely in the Brazil region), and some southeast Asia.There are precipitation increases in eastern Africa, south of the equator, and west and south of South America precipitation decrease.The PCA1-associated SST pattern displays strong signals in the SH mid-latitudes, mostly warming, along with a warming in the Northern Tropical Atlantic and the midlatitude North Pacific (figure 5(f)).Cool SSTs are also displayed in the central tropical Pacific.This pattern, which includes a negative IPO-type pattern and a warming of the tropical North Atlantic, broadly mirrors that of PCA1 in April-September in the tropics.Still, the strong SH signals here are quite distinct.As for the April-September season, the two PC time series of PCA1 (i.e., one for the mean and the other for the 95%) in October-March show strong co-variability, with a correlation coefficient of 0.94 for the first and second modes (figures 5(g) and (h)).

Discussion and conclusions
Utilising an ensemble of historical simulations and two idealized experiments carried out with the ACCESS-ESM1.5,we sought to understand the impact of SST variability on the relationship between decadal mean precipitation and extreme precipitation.The ensemble of historical simulations reproduced the observed strong covariability between decadal mean precipitation and extreme wet precipitation, as in Adamu et al (2022).That is, low decadal mean precipitation is most prominently related to decreases in monthly wet extremes in most of the globe.This is consistent with previous studies, like Zhou and Lau (2017), who argued that fewer wet extremes are likely to occur in regions with negative mean precipitation.
Further, our idealized atmosphere-only simulations showed that in the absence of SST variability, the covariation between the decadal mean precipitation and the wet extremes disappears over a large majority of the globe.This suggests that the major driver of decadal changes in the mean and extremes, and their associated co-variability is driven by internal SST variability.The SST anomalies associated with this co-variability suggests that a PDO/IPO-like structure in the Pacific, and an AMO-like structure in the Atlantic, are the main drivers of decadal-scale co-variability in mean and extreme wet precipitation during April-September.A similar tropical SST structure is also identified from October to March.However, in these months, there is much more weighting on warming in the SH mid-latitude region.
In figure 3, the removal of SST variability in the AGCM SSTA removed experiment leads to substantial changes in the decadal-scale co-variability between mean and extreme precipitation across various regions.It is crucial to note that the weakening of this co-variability does not imply that both mean and extreme precipitation remain static in the absence of SST variability (results not shown).Rather, changes can manifest differently in various areas, with alterations in decadal mean precipitation occurring independently of extreme precipitation, and vice versa.This observation aligns with the idea that SST variability is a significant influencer, but not the sole determinant of variations in mean and extreme precipitation.
In comparison, the covariability between decadal means and extreme dry precipitation is relatively weak in the observations (as shown in Adamu et al 2022) and ACCESS ESM1.5, consistent with Nishant and Sherwood (2021).However, this covariability was still significantly modulated by the removal of SST variability, but over a much smaller portion of the globe (∼40%-45%).The covariability between wet and dry extremes in the ACCESS ESM1.5 and observations are mostly insignificant.Further, the changes to the covariability between decadal-scale wet and dry extremes were largely insignificant when SST variability was removed.
The result from this work demonstrates distinct changes in the covariability of precipitation decadal mean and extremes when SST variability is removed.Consistent with the earlier work of Adamu et al (2022), we hypothesise that decadal changes in the 95th percentile act to alter the decadal mean.Consistent with the strong link between wet extremes and decadal rainfall changes, Adamu et al (2022) find that dry decades are predominantly modulated by changes in positive skewness in monthly precipitation distributions.
Given that here we have shown that the covariability of the mean and wet extremes is modulated by SST variability, and SSTs are known to be relatively slowly evolving, we propose SST information may potentially be used as a longer-term predictor of wet extremes.The ability to carry out such predictions may provide policymakers with the opportunity to properly plan for extreme wet events, which help mitigate the effects of floods, for example (Blenkinsop et al 2021).
Similar experimental designs can be carried out to evaluate the influence of specific oceanic basins on rainfall.For example, what happens when the variability of the tropical Pacific or the tropical Indian Ocean is removed?Also, there could be a benefit from further exploring/evaluating the ability of ACCESS ESM1.5 to reproduce modes of SST variability and their precipitation teleconnections.Our results show that the IPO-like signal is inconsistent with the literature.Further investigation could provide useful information to improve future versions of this model.

Figure 1 .
Figure 1.Correlation coefficients between decadal running means and extremes (a)-(d) and between dry and wet extremes (e)-(f) from the 40-member ACCESS ESM1.5 simulations.The means and extremes are computed in 10 year running windows from 1870-2013.The results for April-September are presented in the left column and October-March in the right column.Note that the correlations are computed for each of the 40 members and the figures show the ensemble mean.

Figure 2 .
Figure 2. of spatial correlations between individual member correlation maps from figure 1, produced from all possible pairings of the 40-member ensemble.The spatial map of each member are correlated against all the other 39 members.Panel (a) shows the distribution of spatial correlations for the April-September season, and Panel (b) is for October-March.

Figure 3 .Figure 4 .
Figure 3.The correlation coefficient between decadal running means and extremes (a)-(d), and between dry and wet extremes (e)-(f) for the AGCM SSTA removed experiment.Regions that are shaded gray have correlation coefficients that do not fall within the ensemble range of the historical ensemble (i.e. the difference is statistically significant).Regions that are white have correlations near zero, defined between −0.1 and +0.1.

Figure 5 .
Figure 5.The first modes of variability computed from PCA for 10 year running means and 95th percentiles for the April-September season (a) and (b), while (c) and (d) displays those for the October-March season.The regression between decadal mean SSTs and PC1 for April-September and October-March from the PCAs are shown in (e) and (f), respectively, with the associated, normalised PC time series shown in (g) and (h) (here we showed for only one ensemble member).Stippled areas in e and f indicate regions that are statistically significantly related to the PCA modes at 95% confidence interval.The correlation coefficient between the mean and 95th percentile PC time series are provided in the titles of (g) and (h).

Figures 5
Figures 5(e) and (f) show the SST regression patterns, calculated by regressing the first PC time series of the decadal-scale wet extremes against decadal SSTs (the regression of the first mode from the decadal mean (not shown) shows a similar pattern given the strong co-variability in figures 5(g) and (h)).The first mode shows strong positive anomalies over the tropical Pacific Ocean, flanked by strong negative anomalies over the northwest Pacific from approximately 20 • N-50 • N.This spatial signature is broadly consistent with the Interdecadal Pacific Oscillation (IPO) and the Pacific Decadal Oscillation (PDO) (Newman et al 2016), both of which are often used nearly interchangeably to describe internally generated Pacific basin decadal SSTs that are proposed to impact rainfall distribution across the globe (Henley et al 2015, Power et al 2021).In the Atlantic Ocean, cooler SST anomalies are identified to the north of the equator that is somewhat consistent with the Atlantic Multidecadal Oscillation (AMO).At the same time, warming is also found on the equator in the east, which is consistent with the Atlantic zonal mode or the Atlantic Nino.Both are also known to have climatic impacts at locations around the globe (Losada et al 2010a, 2010b, 2012, Zampieri et al 2017).