Quantifying the impact of early 21st century volcanic eruptions on global-mean surface temperature

Despite a continuous increase in well-mixed greenhouse gases, the global-mean surface temperature has shown a quasi-stabilization since 1998. This muted warming has been linked to the combined effects of internal climate variability and external forcing. The latter includes the impact of recent increase in the volcanic activity and of solar irradiance changes. Here we used a high-resolution coupled ocean–atmosphere climate model to assess the impact of the recent volcanic eruptions on the Earth's temperature, compared with the low volcanic activity of the early 2000s. Two sets of simulations are performed, one with realistic aerosol optical depth values, and the other with a fixed value of aerosol optical depth corresponding to a period of weak volcanic activity (1998–2002). We conclude that the observed recent increase in the volcanic activity led to a reduced warming trend (from 2003 to 2012) of 0.08 °C in ten years. The induced cooling is stronger during the last five-year period (2008–2012), with an annual global mean cooling of 0.04 °C (+/− 0.04 °C). The cooling is similar in summer (0.05 °C +/− 0.04 °C cooling) than in winter (0.03 °C +/− 0.04 °C cooling), but stronger in the Northern Hemisphere than in the Southern Hemisphere. Although equatorial and Arctic precipitation decreases in summer, the change in precipitation does not indicate robust changes at a local scale. Global heat content variations are found not to be impacted by the recent increase in volcanic activity.


Introduction and motivation
Despite a continuous increase in the concentration of well-mixed greenhouse gases (GHGs), the observed global mean air-surface temperature (GMST) trend has remained more or less steady from 2001 to 2013 (England et al 2014). This rate of temperature change does however not result in a significant slowdown of global warming rate with regard to the 1950-2013 time series (Rajaratnam et al 2015) and the pause was partly attributed to observational errors (Karl et al 2015). The recent observed GMST trend over the period of interest here (2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)  The surface temperature fluctuation can be mainly associated with the internal climate variability through a negative phase of the Pacific Decadal Oscillation (PDO) (Meehl et al 2011, Kosaka and Xie 2013, Trenberth and Fasullo 2013, Guemas et al 2013, England et al 2014and Douville et al 2015. During the last decade, the warming rate has been reduced, along with an increase in heat uptake, mainly in the Pacific and the Atlantic Ocean (Meehl et al 2011, Guemas et al 2013.
External forcing can also cause decadal-timescale changes in GMST. Some studies argue that the solar minimum around 2009 (Kopp and Lean 2011), the recent decrease in stratospheric water vapor concentration (Solomon et al 2010), the increase in emissions of anthropogenic sulfate aerosols (Kaufmann et al 2011, Schmidt et al 2014 and the increase in stratospheric volcanic aerosol emissions (Solomon et al 2011, Fyfe et al 2013, Santer et al 2014, 2015, Haywood et al 2014, Ridley et al 2014, Schmidt et al 2014, Brühl et al 2015, Mills et al 2016 partly explain the reduction of the warming trend. Systematic errors in certain external forcings in CMIP5 simulations of historical climate change explain the difference between simulated and observed warming rates during the 'slowdown' period. CMIP5 climate models are generally forced by unrealistically low stratospheric aerosols optical depth after 2000 (Fyfe et al 2013, among others). The cooling effect of the solar activity was also neglected (Schmidt et al 2014). Biases in simulating El Niño Southern Oscillation phases also lead to GMST overestimation (Schmidt et al 2014). Observational errors (Karl et al 2015) and model-observation comparison biases (Cowtan et al 2015) can also be partly responsible for the apparent gap between models and observations. Richardson et al (2016) have therefore shown that the gap between model simulations and observations are greatly reduced when model outputs are processed the same way as the HadCRUT4 observations over the oceans, and in taking into account the lack of values over the Arctic in the observations. HadCRUT4 is a dataset of observed near-surface air temperature (Morice et al 2012).
Here we study the recent surface temperature fluctuation by analyzing the impact of the volcanic eruptions that have occurred between 2003 and 2012. In order to simulate the impact of the recent volcanic activity on climate we performed two sets of sensitivity experiments, with two different data sets of stratospheric aerosol optical depth (SAOD).
Our approach is similar to that followed by Solomon et al (2011), Fyfe et al (2013, Ridley et al (2014) and Haywood et al (2014), who analyzed the impact of the recent volcanic eruptions on the GMST. The main objective of Fyfe et al 2013 was to explore the consequences of systematic errors in model representation of early 21st century volcanic aerosol forcing (using a near-zero radiative forcing as a baseline). Here we prefer using an alternative approach with an SAOD baseline set at a fixed value, deduced from a weak volcanic activity period (1998)(1999)(2000)(2001)(2002). We address the impact of the recent increase in the volcanic activity, in comparison to a period of a low-volcanic activity as a baseline. We mainly assess the impact of the moderate strengthening in the volcanic activity on a recent decade (2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012) rather than documenting the impact of using biased low forcing. To analyze the impact of natural volcanic activity variability in (the absence of strong volcanic events) we have to deal with weaker differences in SAOD, than in previous studies (Fyfe et al 2013, among others). The impact of such a low change in radiative forcing is thus still an open question. The question of the misrepresentation of the volcanic external forcing in global climate models is therefore indirectly addressed.
Our scientific question: What is the impact, on the GMST, of the recent of moderate volcanic activity (2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)  The coupled model is initialized with the GLORYS2V1 ocean reanalysis (Ferry et al 2012) and the atmosphere initial state is given by a sea-surface temperature (SST)-forced ARPEGE simulation. The simulations are full-field initialized in 2002 (in November). We performed six realizations, which differ by their atmospheric initial state, corresponding to different days in November. All simulations are forced by the estimated historical values in GHGs, solar irradiance, black carbon, particulate organic matter, dust, sea salt and sulfate aerosols. The optical depths of tropospheric aerosols are given by Szopa et al (2013).
The ability of CERFACS-HR to simulate climate is provided in MO17. Before the drift correction, simulations exhibit large positive biases of SST over the Austral Ocean and over the equatorial Pacific and Atlantic Ocean (up to 4°C), the Arctic experiences a cold bias of 5°C. These biases are commonly obtained with climate model simulations. After drift correction CERFACS-HR is able to simulate and to predict the global mean surface temperature for the upcoming five years and exhibits high skill in predicting SST over the North Atlantic, equatorial Atlantic and the Indian Oceans (MO17). CERFACS-HR has also high skill in predicting decadal trends in Arctic sea-ice extent and volume. Moreover CERFACS-HR has been used in a previous study to perform a suite of decadal hindcasts Environ. Res. Lett. 12 (2017) 054010 (10 year hindcasts starting each year from November 1993 to November 2012) that reproduce a reduced surface warming rate during the recent decades (MO17). In these hindcast simulations, the 2003-2012 trend in GMST is associated with a negative phase of the PDO, and is therefore mainly explained by internal climate variability. This result is likely to be linked to the model initial state, since the PDO and the Interdecadal Pacific Oscillation (IPO) shifted from a positive to a negative phase in the early 2000s Fasullo 2013, England et al 2014). This issue is discussed later, in section 3.4. However, MO17 carried an important limitation, i.e. an unrealistic evolution of the volcanic eruptions due to the use of a background condition in SAOD instead of realistic SAOD data. Here we analyze the added value of a realistic representation of the volcanic activity.
To meet this objective, two sets of simulations are performed: The first set, hereafter referred as VER, includes six members using the SAOD data of Vernier fluxes (shortwave þ longwave) (Wm À2 by decade). The circles represent each BAK and VER simulations (six members), the cross represents the ensemble-mean and the triangle the ensemble-median. The gray shading and the discontinuous lines are used to delimit the spread (þ/À one standard deviation around the ensemble mean) with respectively the BAK and VER ensemble. By convention negative values are used for upward fluxes, positive values are used for downward fluxes. A negative trend is thus associated with a cooling of the troposphere while a positive trend is associated with a warming of the troposphere.
Environ. Res. Lett. 12 (2017) 054010 allows quantifying the impact of the recent volcanic activity in comparison with a period of minimal volcanic activity. Here we do not document the consequences on the climate simulations of using a biased low volcanic forcing (with a near-zero value in SAOD), as done in many studies (Solomon et al 2011, Fyfe et al 2013, Ridley et al 2014. The background in SAOD is however never equal to zero (as highlighted in Solomon et al 2011) and our choice of the baseline value of SAOD is more adapted to our objective.
The impact of the recent volcanic eruptions is quantified by computing the VER-BAK differences.
Climate models do not perfectly simulate the observed climate. When initialized with observations models drift toward their preferred imperfect climatology (Meehl et al 2009). This drift leads to biases in hindcasts and has to be removed. We used an additional set of ten 10 year hindcasts, initialized every year from 1993 to 2002 (5 members Â 10 start dates), described in MO17, to estimate a drift correction, following the World Climate Research Program recommendations (ICPO 2011). The drift was then removed from VER and BAK at each leadtime (time step after the model initialization) (see supplementary materials available at stacks.iop.org/ ERL/12/054010/mmedia for further details on the bias adjustment method). Assuming that the drift is equal in the two sets of simulations, we did not remove the drift prior to computing the VER-BAK differences (in what follows, we only remove the drift when showing separately VER and BAK results). The VER-BAK differences are thus computed at the CERFACS-HR horizontal resolution. We checked that the result is not dependent on the removal of the drift prior to compute the anomalies (not shown).

Data
We compare the simulated GMST to several data sets of near-surface air temperature:

Radiative forcing
The 10 year trends of the top of the atmosphere (TOA) net shortwave, longwave and net radiative fluxes are shown in figures 1(b)-(d). The inter-member spread is large in both BAK and VER simulations. We found a large interannual variability in TOA for both BAK, VER, we assume that this can lead to large intermember spread over short-periods (ten years). As the six-simulation ensemble average is strongly influenced by outliers, we also show the ensemble median. Both BAK and VER exhibit negative 10 year trends in TOA net shortwave ( figure 1(b)). The VER trend is stronger than the BAK trend, indicating a stronger decrease in the net shortwave radiation. Moreover the VER-BAK difference is greater than 1 BAK standard deviation (the standard deviation obtained from the six realizations of BAK, figure 1(b)). We split the net shortwave flux into its incoming and outgoing parts (not shown). As the former is the same in both simulations, this indicates that recent eruptions increased the outgoing shortwave radiation, due to the volcanic cloud albedo effect. Santer et al ( 1  (c)). This denotes a stronger absorption of upward longwave fluxes in VER, which lead to a warming of the troposphere. It is not strong since the differences in the mean and the median are weaker than 1 BAK standard deviation. The VER minus BAK difference in TOA longwave is due to volcanic eruptions that inject sulphur dioxide (and to a lesser extent, hydrogen sulfide) into the stratosphere, which are converted into sulfate aerosols that act to both scatter solar radiation and absorb longwave and near-infra-red radiation, as shown in Robock (2000) and Timmreck (2012).
Finally the VER simulations exhibit a more negative (i.e. upward) trend in net TOA heat fluxes Environ. Res. Lett. 12 (2017) 054010 than BAK and the difference between VER-BAK is even more robust when considering the medians (figure 1(d); online supplementary table S1). This difference indicates a stronger increase in outgoing heat fluxes in VER compared to BAK. The difference, of 0.21 W m À2 in ten years (median) is consistent with the modelled studies of Brühl et al (2015) who obtained a decrease in the radiative forcing by up to 0.2 W m À2 due to the volcanic eruptions between 2005 and 2011.
It is therefore expected that VER simulates a colder mean climate than BAK, which is consistent with the known impact of volcanic eruptions on radiative fluxes at the top of Earth's atmosphere.

Global mean surface temperature
The 2003-2012 GMST trend ranges from a weak warming in BEST and in Cowtan and Way to a weak cooling in ERAI, in MLOSTand in GISTEMP (table 1).
Observed 10 year trends have large uncertainties for all data sets (third column of the table 1). The observed trends are close to the 1998-2012 trend (0.03°C per decade), a result found by Kosaka and Xie (2015) with the HadCRUT 4 dataset.
The observed data sets have different horizontal resolutions and use various infilling and coverage over the polar regions. Hansen et al (2010) and Cowtan and Way (2014) have shown that discrepancies across the datasets in GMST are mainly due to the estimated temperature in polar regions, where observations are sparse.
The BAK 2003-2012 trend is warmer than the corresponding ERAI trend, with a warming of þ0.12°C (þ/À 0.05°C) in ten years (close to the observed GMST 1951-2012 trend, of þ0.11°C by decade, according to Kosaka and Xie 2015) (table 1 and figure 2(a)). In VER, the 2003-2012 trend is also positive (þ0.04°C þ/À 0.07°C per decade). This trend is close to BEST and 'Cowtan and Way' , which are the most reliable observations for the analysis of the GMST, due to a better representation of the surface temperature at high latitudes (Dodd et al 2015). The difference between VER and BAK trend is determined from the time series of paired differences between VER and BAK. It reduces noise levels by subtracting variability components common to VER and BAK. The BAK-VER trend is statistically significant at the 95% confidence level according to a Spearman's rank correlation test. This suggests that the recent volcanism activity has a significant impact on the GMST, reducing the warming trend by about 0.08°C in ten years, compared to a period of weak volcanic activity.
BAK and VER GMST differences get stronger with time ( figure 2(a)). This can be due to the increase in SAOD after 2007 with stronger peaks mainly due to the Tavurvur  In figure 2 we chose ERAI as our observation data set since it provides information for a large set of variables (see the additional analyses in the supplementary materials) with a complete Earth coverage (avoiding model-observation differences due to different spatial coverages, such as missing values over the Arctic as reported in Hansen et al 2010 andWay 2014). Note that ERAI can be considered as a reference, since estimates of Arctic temperatures and temperature trends are realistic (Simmons et al 2014, Simmons andPoli 2015). In figures 2 (b)-(d), we show the VER-BAK GMST differences for the 5 last years of the simulations (2008)(2009)(2010)(2011)(2012), when the difference between the VER and BAK AOD files is the strongest and when the observed impact of the volcanic activity is the strongest since 2003 (Santer et al 2015). The global impact is of -0.04°C (þ/À 0.04°C). The difference is similar over land (À0.05°C þ/À 0.06°C) than over the oceans (À0.04°C þ/À 0.03°C) (figures 2(c),(d)). All these differences are statistically significant at a 95% confidence level following a Monte Carlo approach (more details are available in the supplementary materials; table S1 and table S2). VER simulates a colder mean climate than BAK for the global Earth ( figure 2(b)), the land (figure 2(c)) and the ocean ( figure 2(d)).
The ERAI-VER difference is however stronger than the VER-BAK difference (and stronger than 1 VER standard deviation on global average). There are clearly other phenomena to take into account to reproduce the observed slightly negative trend, such as, for instance, the cooling over Eurasia in winter, which arises essentially from atmospheric internal variability (Li et al 2015). Table 1. 10 year GMST trend for the different observations, ERAI and the modeled 10 year GMST trend. The 2003-2012 GMST trend error is given by half the 2.5%-97.5% regression coefficient confidence interval. For VER and BAK, the inter-member spread in GMST trend is added. 3.3. Regional response in temperature, precipitation and heat content Haywood et al (2014) have shown that the impact of the recent volcanic eruptions is not spatially homogeneous, for example stronger in the Northern Hemisphere than in the Southern Hemisphere. In this section, we focus on the regional responses in temperature and precipitation. Moreover we explore the seasonality of the response. From November to April (NDJFMA, which includes the boreal winter) the difference in GMST is of À0.03°C (þ/À 0.04°C). The impact is significant only over several restricted areas ( figure 3(a)). Precipitation change only exhibits sparse significant anomalies ( figure 3(c)).
Impacts are of comparable magnitude from May to October (MJJASO, which encompasses the boreal summer) with a decrease in GMST of 0.05°C (þ/À 0.04°C). The zonally-averaged temperature decreases over the northern hemisphere, due to a cooling of the Arctic. The subpolar gyre warms and the tropical Atlantic and eastern Pacific ocean cool down ( figure 3(b)). A stronger impact over the Arctic is consistent with Fyfe et al (2013) and Haywood et al (2014). Moreover the cooling over the Arctic is stronger in MJJASO than in NDJFMA, since it receives more solar flux in summer than in winter, allowing a more efficient direct effect of the volcanic forcing. Precipitation decreases over the equator and south of it, as also suggested in Fyfe et al (2013). This result however disagrees with Haywood et al (2014), who have found a southward shift of the tropical precipitation due to a cooling of the subtropical Atlantic Ocean. The precipitation pattern consists in an increase in precipitation over the western Pacific Ocean and a decrease over the western coast of South Environ. Res. Lett. 12 (2017) 054010 America, along the equator, which is consistent with a La Niña pattern. It has been shown that major volcanic eruptions can affect the ENSO variability, leading to an El Niño (Ohba et al 2013) or a La Niña (Maher et al 2015) event several months after the eruption. Here there is no robust change in ENSO variability (not shown).
The volcanic eruptions can also impact the North Atlantic Oscillation (NAO) variability (Shindell et al 2004). We obtain a tripole in sea-level pressure (SLP) anomaly: a positive pressure anomaly north of Europe and negative ones west of Spain and over northeastern Canada, i.e. that project on a negative phase of NAO (online supplementary figure S1). The negative NAO pattern is consistent with a decrease in surface zonal wind speed over the North Atlantic (figure S1), and the warming of the north Atlantic SST. However, a NAO index, derived from our simulations, does not indicate a robust change in the NAO variability (not shown). Gleckler et al (2016) argued that the recent volcanic eruptions led to a decrease in ocean heatuptake. We however found no strong differences between BAK and VER when considering the global mean heat-content surface to 700 m, 700 m to 2000 m and the deep ocean (>2000 m) (online supplementary figure S2). Here the heat uptake increase is weak and only located to the eastern equatorial Pacific Ocean (online supplementary figure S3).
3.4. Impact of the initial condition and mean state Pohlmann et al (2004) found that the North Atlantic SSTs, Nordic Seas and Southern Ocean exhibit predictability on multidecadal time scales, owing to the model initialization and the oceanic inertia. Numerous predictability studies have highlighted the importance of the initialization to predict the GMST (Bellucci et al 2013, García-Serrano et al 2015, Karspeck et al 2015. Here we analyze the 10 year trend in GMST for both BAK and VER. In both simulations, temperature increases over the subtropical Pacific Ocean and decreases over the eastern Pacific Ocean, exhibiting a negative phase of the PDO (online supplementary figure S4). CERFACS-HR systematically reproduces a negative phase of the PDO over the 2003-2012 period, as also shown in MO17 due to the ocean initialization (these simulations have common ocean initial conditions).
On the one hand the predictability of the SSTs is mainly given by the model initialization. On the other hand the mean state of the climate system may modulate the impact of the volcanic eruptions (Zanchettin et al 2013). We can thus hypothesize that the impact of the volcanic eruptions could be different without the simulated negative phase of the PDO, and the associated abnormally low GMST.   (2008)(2009)(2010)(2011)(2012) in (first row) temperature (in°C) and (second row) precipitation (mm.day À1 ). Dots indicate that anomalies of precipitation (air surface temperature) are significant at the 90% (95%) confidence level according to a Student's t-test. The side plot represents the zonallyaveraged VER-BAK differences in temperature and precipitation. Blue filled-circles represent the anomalies of precipitation (air surface temperature) considered significant at the 90% (95%) confidence level according to a Student's t-test.

Discussion/conclusion
Environ. Res. Lett. 12 (2017) Santer et al (2014Santer et al ( , 2015, Brühl et al (2015) and Mills et al (2016) that the recent volcanic activity has reduced the GMST and is therefore one of the causes of the recent reduced warming trend.
A cooling of 0.04°C (þ/À 0.04°C) over the 2008-2012 period was found in this study. Such result is consistent with Haywood et al (2014), in which the decrease of temperature is of 0.02°C-0.03°C. A stronger response found in the Northern Hemisphere than the Southern Hemisphere is also consistent with this author.
Both BAK and VER simulations exhibit a negative phase of the PDO, which is therefore not the only mechanism responsible for a global warming rate slowdown. For instance, even limited increase in volcanism activity, as it occurred in the last decade, has also exacerbated the eastern Pacific cooling.
The induced cooling (0.04°C) is weaker than in Solomon et al (2011)  Here we used a baseline based on a weak volcanic activity (a more realistic value since the volcanic activity is not expected to stop) and are dealing with weaker anomalies in radiative forcing. The obtained cooling is thus weaker than in the aforementioned studies. We however assume that this experimental protocol is noteworthy since we can document the impact of natural volcanic activity variability, in absence of intense volcanic events, rather than documenting the impact of using biased low forcing.
Vernier et al (2011) SAOD neglects substantial amounts of volcanic aerosols below 15 km and therefore underestimates total radiative forcing due to the recent eruptions (Ridley et al 2014 andMills et al 2016). The impact of including the recent volcanic eruptions in climate models could thus have been stronger than previously highlighted and obtained here. Ridley et al (2014) estimated a global cooling of 0.05 to 0.12°C since 2000 with a better estimation of the global volcanic aerosol forcing.
Results obtained in this study show that moderate volcanic eruptions cause a small cooling of the Earth and highlight the importance of taking into account a realistic volcanic forcing in the climate models. However, this volcano-related cooling is not sufficient to fully explain the recent surface temperature stagnation. The cooling generated by volcanic eruptions is thus not the unique cause for the recent decrease in the overall trend of temperature increase.
The impact of the recent volcanic eruptions is a promising topic for understanding the GMST evolution and its better understanding can help to improve climate model simulations since the stratospheric AOD are not well represented by the CMIP5 climate models in the recent decades (Fyfe et al 2013). Such a study should however be extended to a multimodel analysis, in order to assess the robustness of the results with regard to model sensitivity, and/or with different ocean initial conditions.