Correction notice Nature Climate Change 2, 524 549 (2012) Human-induced global ocean warming on multidecadal timescales P. J. Gleckler, B. D. Santer, C. M. Domingues, D.W. Pierce, T. P. Barnett, J. A. Church, K. E. Taylor, K. M. AchutaRao, T. P. Boyer, M. Ishii and P. M. Caldwell In the version of this Supplementary Information file originally posted online, some figures were missing. The errors have been corrected in this file 10 July 2012.
SUPPLEMENTARY INFORMATION DOI: 10.1038/NCLIMATE1553 Human-induced global ocean warming on multidecadal timescales Gleckler et al. Volume average temperature Ocean Heat Content (OHC) is a commonly used measure of ocean temperature/heat changes within a vertical column (e.g. 0-700 m). However, if OHC is to be calculated from measurements, the sparseness of historical in-situ data necessitates use of an objective method to infill in areas were measurements are not available. This is a critical part of the work done by groups that make available near-global estimates of upper ocean temperature changes. An alternative approach, used in previous D&A studies and continued here, is to work with Volume Average Temperature (ΔT) rather than OHC. With ΔT, it is possible to focus only on when and where measurements were made, and to compare findings with estimates where infilling is performed. This is important for D&A work because it is known that estimates of noise can be quite sensitive to sampling. Observational Data We make use of five publically available in-situ based temperature estimates (Table S1). Three of them have become available relatively recently and include XBT fall rate bias 1-3 corrections. The two older estimates pre-date bias corrections and thus enable us to evaluate the impact of the corrections on our analysis. Our NATURE CLIMATE CHANGE www.nature.com/natureclimatechange 1
subsampling tests (described below) are only performed with the Levitus and Ishii data because the Domingues et al. (2008) reconstruction is only available for two depth intervals (0-300m, 300-700m). Table S1: Sources used for temperature estimates Reference ID Period XBT Bias correction Domingues et al. (2008) Domingues08 1950-2003 Y Ishii et al. (2006) Ishii06 1950-1998 N Ishii et al. (2009) Ishii09 1945-2006 Y Levitus et al. (2005) Levitus05 1955-1998 N Levitus et al. (2009) Levitus09 1955-2008 Y Model Simulations Historically forced (20CEN) and pre-industrial control (PICNTRL) simulations The model simulations examined here were contributed to the World Climate Research Programme s Coupled Model Intercomparison Project Phase 3 (WCRP CMIP3) data archive. Our analysis includes models for which all necessary data was available, namely upper ocean temperature for both a pre-industrial control run (PICNTRL) and at least one simulation of the 20th Century (20CEN), which includes time varying external forcings. This is the same collection of CMIP3 models used in earlier studies 1,4,5. Table S2 identifies the modeling groups that provided sufficient model output for us to perform our analysis. Several groups provided results for more than one model
configuration, and many of them contributed multiple realizations of the 20CEN experiment (see Table S3). Details of the natural and anthropogenic forcings used in the CMIP3 20CEN simulations are summarized in Table S4. TABLE S2: CMIP3 Pre-industrial control simulations (PICNTRL) CENTER MODEL ID Length (yrs) Canadian Centre for Climate Modeling and Analysis CCCma-CGCM3.1 1000 (CAN) Meteo-France/Centre National de Recherches CNRM-CM3 500 Meteorologique (FRANCE) Commonwealth Scientific and Industrial Research CSIRO-Mk3.0 380 Organization (AUSTALIA) Geophysical Fluid Dynamics Laboratory (U.S.) GFDL-CM2.0 500 Goddard Institute for Space Studies (U.S.) GISS-AOM 250 / 250 a Goddard Institute for Space Studies (U.S.) GISS-EH 400 Goddard Institute for Space Studies (U.S.) GISS-ER 500 Institute for Atmospheric Physics (CHINA) FGOALS-g1.0 350 Meteorological Research Institute (JAPAN) MRI-CGCM2.3.2 350 Center for Climate System Research, National Institute MIROC- 500 / 100 b for Environmental Studies, and Frontier Research Center for Global Change (JAPAN) CGCM2.3.2 (med / high res) National Center for Atmospheric Research (U.S.) CCSM3_0 150 / 500 a Hadley Centre for Climate Prediction and UKMO-HadCM3 340 Research(U.K.) a Two PICNTRL runs exist for this model. b Two versions of this model, one at med resolution and the other at high resolution
TABLE S3: Historically forced simulations (20C3M) MODEL ID Ensemble Spawn date a size CCCma-CGCM3.1 5 1850 b CNRM-CM3 1 2040 CSIRO-Mk3.0 3 1871, 1881, 1891 GFDL-CM2.0 1 1 GISS-AOM 2 1850, 1850 c GISS-EH 4 2000, 2010, 2020, 2030 GISS-ER 7 FGOALS-g1.0 3 1850, 1855, 1860 MRI-CGCM2.3.2 5 MIROC-CGCM2.3.2 3 / 1 2300, 2400, 2500 / 1 (med / high res) CCSM3_0 6 g 360, 380, 380 e,c, 410 e,c, 460 e,c,540 e,c UKMO-HadCM3 2 1860, 1960 a base-time, which is arbitrary, is handled differently by each group and corresponds to times in the official CMIP3 database b Each realization uses same initial condition as PICNTRL, with a small perturbation added c Spawned from PICNTRL2 g usable forced runs 1, 2,6,7,8,9; contemporaneous PICTNRL data not available for other runs TABLE S4: 20C3M external forcing MODEL ID G O SD SI BC OC MD SS LU SO V CCSM3_0 Y Y Y - Y Y - - - Y Y GFDL-CM2.0 Y Y Y - Y Y - - Y Y Y GISS-EH Y Y Y Y Y Y Y Y Y Y Y GISS-ER Y Y Y Y Y Y Y Y Y Y Y MIROC-CGCM2.3.2 1 Y Y Y Y Y Y Y Y Y Y Y MRI-CGCM2.3.2 Y - Y - Y - - - - Y Y CCCma-CGCM3.1 Y - Y - - - - - - - - CNRM-CM3 Y? Y??????? - CSIRO-Mk3.0 Y Y Y -?????? - GISS-AOM Y Y Y - - - - Y - - - FGOALS-g1.0 Y Y Y? - - - - - - - UKMO-HadCM3 Y Y Y Y Y Y - - Y Y - 1 valid for med and high resolution G= Well-mixed greenhouse gases, O= tropospheric and stratospheric ozone, SD = Sulfate aerosol direct effects, SI=Sulfate aerosol indirect effects, BC = Black Carbon, OC = Organic carbon, MD = Mineral dust, SS = Sea salt, LU = Land use change, SO = Solar irradiance, V = Volcanic aerosols
SRESA1B Scenario Simulations The multi-model V and NoV historically forced results in Figure 1C have been extended by 10 years with available SRESA1B future scenario simulations. There is a small discontinuity in year 2000 in the V and NoV curves of Figure 1C because not all of the historical runs were used to initialize SRESA1B simulations. With respect to the 20CEN number of realizations listed in Table S3 (above) for each model, the following number (#) of SRESA1B simulations are missing for each model: csiro_mk3_0 (2), giss_model_e_h (1), giss_model_e_r (3), ncar_ccsm3_0 (3) and ukmo_hadcm3 (1). In addition to being based on a few number of realizations that the 20CEN multi-model averages, 20CEN-SRESA1B forcing discontinuities are know to exist in some models. Processing of Model Simulations Calculation of three dimensional anomaly data on a common grid To perform our subsampling analysis described below, it is necessary to have ocean temperature model output on the same (Latitude x Longitude x Level) grid as the observations. We make use of the same annual mean model data transformed to this common grid as used in previous studies 5. Annual mean anomalies were then computed for each PICNTRL simulation with respect to the local time mean. For the 20CEN simulations, late 20 th Century anomalies were computed with respect to the 1957-1990 local climatology, consistent with the period used in
earlier D&A OHC studies 6-8. All available data is used in these PICNTRL and 20CEN anomaly time series, i.e., they are infilled (IF). Subsampling (SS) Sampling model data only when and where historical measurements have been made enables the ΔT of the two to be compared in a consistent manner. The PICNTRL anomaly data are used to produce sub-sampled 4D time series for each PICNTRL simulation as follows: We apply the Ishii and Kimoto (2008) time varying mask of measurement availability (1961-2007) in a repetitive non-overlapping fashion to each spatial complete 4D PICNTRL anomaly time series. (1957-1960 data are not used because of the paucity of data). This Ishii mask identifies locations (in latitude, longitude and depth) were there was at least one measurement made during each year. Model data is retained at locations and times where the mask identifies at least one measurement, and elsewhere the 4D field is set to missing (with zero volume). The ocean volume in each basin is thus a function of time, which is why our analysis is performed with volume average temperatures rather total ocean heat content. The measurement mask of 1961 is applied to the first year of a PICNTRL simulation, 1962 to the 2 nd year, and so on. This procedure of subsampling the model output with the data mask begins to repeat at year 48 of the PICNTRL and continues repeating until the end of the control run is reached. We have truncated the residual years at the end of the subsampling process so that the record length for each model is divisible by the length of the number of years in the observational mask. Any impact of discontinuities associated with the end points of
the time varying mask (e.g., between year 47 and 48) will increase our noise estimate and have a conservative impact on our D&A analysis, i.e., they make detection more difficult. Spatially complete and sub-sampled 4D data are also produced for each 20CEN anomalies time series with respect to 1957-1990. The model output are subsampled with the 1961-1999 Ishii mask to produce sub-sampled versions of the 20CEN simulations. Many CMIP3 20CEN simulations end in 1999, which we use as our common end-date for our multi-model ensembles (described below). The 20CEN simulations are initialized from a state of the corresponding PICTNRL (see Table S2), with start dates ranging from 1850 to 1900. Our focus is on the late 20 th Century of the forced runs, from 1960-1999. Basin volume average temperatures We use the 4D spatially complete and sub-sampled temperature anomalies for each PICNTRL and 20CEN simulation to compute 0-700 m volume average temperature time series for the NAtl, SAtl, NPac, SPac, NInd, Sind, and global (glb) basins defined in the earlier works 6-8. ΔT is computed for all spatially complete and subsampled data. Estimating control run drift and noise For each PICNTRL, we estimate the long-term drift associated with model spin up. We do so in each basin using simple quadratic (Q) and cubic (C) fits. A linear fit is
not sufficient for many simulations, and higher order fits risk removing longer-term variability that is not associated with the simulation drift. Simulation drift in each basin is removed from the original PICNTRL ΔT time series to provide estimates of basin-scale natural variability or noise. These data are important to the detection and attribution analysis discussed below. Drift removal from forced simulations All of the forced runs we use have contemporaneous data available (See Table S2) from the PICNTRL simulation that the forced run was initiated from. One method that has been used to remove simulation drift from forced simulations is to do a straight subtraction of the contemporaneous control from the forced run (e.g., AchutaRao et al., 2007). This risks inflating the variance since the higher frequency variability of the two is combined. We remove the drift from each 20CEN simulation basin by subtracting the corresponding segment of the control drift (i.e., the cubic or quadratic fit) contemporaneously from each 20CEN simulation. Anomalies are taken in our reference frame with respect to 1957-1990. We have found the drift structure in the CMIP3 control runs to be highly model dependent and spatially complex. Multi-model ensembles Previous studies have demonstrated the importance of volcanic forcing in the CMIP3 simulations. To isolate these impacts, we perform our multi-model analysis on subsets that include the volcanic eruptions (V) and those do not (NoV). Our
multi-model noise estimates are obtained by "pooling" the PICNTRL basin-scale noise estimates (i.e., with drift removed) of individual models. The resulting concatenated global ocean time series is shown for both the V and NoV model subsets in Fig. S2. Fingerprint Analysis Most D&A studies use models to estimate the pattern of response (i.e., the fingerprint) to human-caused changes in greenhouse gases and other forcings. The first Empiral Orthogonal Function (EOF) of the signal is often used as an estimate of the model fingerprint when the signature of interest is present. In the multi-model approach, an Ensemble Common Response (ECR) to external forcing is obtained by averaging the 20CEN runs together. In practice, if more than one realization of the 20CEN experiment is available for a single model (differing only in their initial conditions), these individual realizations are first averaged together before averaging across other models. Our ECR fingerprints are estimated from the first EOF of the 1960-1999 structural evolution of temperature in our six ocean basins. We utilize a methodology 12 that generalizes a fixed pattern "fingerprint" used in single-model studies, and is a complimentary approach to previous OHC D&A studies 6-8 that rely on a space-time sequence of spatial patterns representing the evolution of the signal.
Method for Estimating Detection Time. Our analysis is based on the original data without performing any optimization. To define detection times, observations and model-based noise estimates are projected onto the fingerprint yielding (respectively) a test statistic time series Z(t) and a signal-free time series N(t). We fit least-squares linear trends of increasing length L to Z(t) and then compare these with the standard error of the distribution of non-overlapping L-length trends in N(t). Detection is stipulated to occur when the trend in Z(t) exceeds and remains above the 5% significance level. The test is one-tailed, and we assume a Gaussian distribution of trends in N(t).
REFERENCES 1 Domingues, C. et al. Improved estimates of upper-ocean warming and multidecadal sea-level rise. Nature, 1090-U1096, doi:doi 10.1038/nature07080 (2008). 2 Levitus, S. et al. Global ocean heat content 1955-2008 in light of recently revealed instrumentation problems. Geophysical Research Letters, -, doi:artn L07608 DOI 10.1029/2008GL037155 (2009). 3 Ishii, M. & Kimoto, M. Reevaluation of historical ocean heat content variations with time-varying XBT and MBT depth bias corrections. Journal of Oceanography, 287-299, doi:doi 10.1007/s10872-009-0027-7 (2009). 4 Gleckler, P. et al. Krakatoa lives: The effect of volcanic eruptions on ocean heat content and thermal expansion. Geophysical Research Letters, -, doi:artn L17702 DOI 10.1029/2006GL026771 (2006). 5 AchutaRao, K. M. et al. Simulated and observed variability in ocean temperature and heat content. Proceedings of the National Academy of Sciences of the United States of America 104, 10768-10773 (2007). 6 Barnett, T., Pierce, D. & Schnur, R. Detection of anthropogenic climate change in the world's oceans. Science, 270-274 (2001). 7 Barnett, T. et al. Penetration of human-induced warming into the world's oceans. Science, 284-287, doi:doi 10.1126/science.1112418 (2005). 8 Pierce, D. W. et al. Anthropogenic warming of the oceans: Observations and model results. Journal of Climate 19, 1873-1900 (2006).
Supplemental Figures NCLIM 11110929 Figure S1. As for Figure 1, but for basin scale ΔT results.
Figure S2. Number of in situ observed temperature profiles per year for 0 300 m (black) and 300 700 m (gray), based on measurements from bottles, XBTs, low and high resolution CTDs, and Argo floats.
Figure S3. Global mean volume average temperature (0 700m) anomaly time series from available pre industrial control run simulations ( C). Control runs were pooled for: A) NoV models and B) V models. In both cases, results are for spatially complete data with quadratic drift removal.
Figure S4. Same as Figure 5C in main text except all results are based on: A) cubic and quadratic drift removal; and B) infilled and sub sampled data. Results are only shown for the newer, corrected Ishii and Levitus ΔT estimates.
Figure S5. Same as Figure 5, except using a start date of 1960 (instead of 1970).
Figure S6. As for Figure 5C in main text, except that results shown here compare S/N ratios obtained with V and NoV fingerprints.