Statistical and Dynamical Approaches to Probabilistic Decadal Climate Prediction

Ana Maria Kupresanin (14-ERD-095)

Abstract

The growing societal interest in climate change has generated a need to understand near-term climate predictions that can assist in the decision-making affecting time scales of about a decade. We developed a statistical model to assess the uncertainty in projections of the climate in the future. The statistical model blends different sources of information from the database of multi-model ensembles of global climate model simulations and observational data. We examined the differences between satellite- and model-based estimates of tropospheric temperature change. We made a contribution to scientific explanations for the post-1998 “hiatus” in surface and tropospheric warming by showing that signals of late 20th and early 21st century volcanic eruptions are statistically detectable in several climate variables. The work has so far contributed to two scientific publications.

This project has only scratched the surface of climate science problems to which statisticians can contribute. The project has identified the need for the Laboratory to establish an interdisciplinary research program around climate. This program should address the critical need for training a new generation of interdisciplinary researchers who can address challenging scientific problems that require complex data analysis by developing and using the necessary sophisticated statistical methods. This project brought two new hires and two summer students to the Laboratory, and established academic collaborations.

Background and Research Objectives

The current state of the art in providing robust global climate projections for the 21st century is based on mining the Coupled Model Intercomparison Project 5 (CMIP5) database, which is a multi-model ensemble of global climate model simulations. This fact is reflected in the latest report by the Intergovernmental Panel on Climate Change. The core set of simulations in the CMIP5 that are used to address 21st century climate change are initialized in 1850 and simulate the 20th century climate using observed boundary conditions (i.e., atmospheric composition, including greenhouse gases, solar forcing, land use, etc.), and then simulate the 21st century under different anthropogenic greenhouse gas scenarios. The variation across global climate model simulations, when averaged over any given decade in the late 21st century, is driven not only by the emission scenario and the structural differences in the models (model uncertainty), but also by the timing (the mode) of natural internal climate variability (e.g., various oscillations in the ocean). However, for the next few decades (10–30 years horizon), the climate will be greatly influenced by the exact amplitude and timing of these various natural internal variability phenomena. Furthermore, this is the time horizon that is of most interest to governments and businesses, and in general for near-term social impact and regional risk assessment of climate change

To respond to this need, the latest database CMIP5 includes a new set of simulation experiments aimed at providing decadal climate simulations.1 The current understanding of the decadal prediction accuracy of global climate models is still in its early stage.2,3 This understanding includes how to assess the accuracy of such simulations, their biases, and in particular how to leverage a multi-model ensemble of such simulations to provide more accurate decadal climate projections.

Decadal predictions are intended to address the evolution of the climate over periods of one or two decades as the climate responds to the internal variability of the system as well as the external forcings.2 For time horizons of a decade or so, climate forecasts are dominated by the phase of low-frequency oscillations in the oceans, the trend of external forcings at the time of initialization, and forcings due to continued emissions of greenhouse gases. Decadal predictions are obtained by initializing a global climate model with the current state of the system. The initialization method varies depending on the modeling center. After some initial adjustments, the model represents the initial state of the real system and evolves synchronously with it. The prediction of climate variability at this time horizon enhances the ability to prepare for heat waves or droughts, changes in fisheries regimes, or increased hurricane activity, to mention a few of the many possible consequences of climate change that can have important societal impact. Decadal predictions are different from the long-term, un-initialized predictions that have been the traditional focus of simulations of long-term climate change. The main objective of our project was to learn these uncertainties statistically from a multi-model ensemble of decadal simulations of the climate in the past and using that knowledge to produce probabilistic predictions 10–30 years out.

Our work was centered on blending information from long simulations of global climate models, decadal simulations, and observational data. In conjunction with the probabilistic predictions of future climate and blending information from observations and simulations, two discrepancies often arise, namely (1) the disagreement of observational data and climate model simulations and (2) competing explanations for warming “hiatus.” In addition to developing a statistical framework for probabilistic decadal climate predictions we also addressed these two questions.

There are several important observations to make before discussing the disagreement of observational data and climate model simulations. Reliable thermometer measurements of large-scale changes in Earth’s surface temperature are available for more than a century, and document warming of roughly 0.85ºC since 1880, with the three warmest decades in the most recent portion of the record.4 Satellite-based estimates of trends in tropospheric temperature cover a shorter period of time (late 1978 to the present), but also provide independent confirmation of planetary-scale warming.5, 6, 7 Although observational and model temperature data provide compelling evidence for the existence of a “discernible human influence” on global climate,4, 8 studies of temperature change continue to yield interesting and important scientific puzzles. Examples include apparent differences between surface and tropospheric warming rates in observational records,9 and differences between modeled and observed warming trends.10, 11, 12

We examined the differences between satellite- and model-based estimates of tropospheric temperature change. We assessed the validity of two highly-publicized claims: that modeled tropospheric warming is a factor of three to four larger than in satellite and radiosonde observations,5 and that satellite tropospheric temperature data show no statistically significant warming over the last 18 years.

A number of scientific explanations have been advanced for the post-1998 “hiatus” in surface and tropospheric warming. One view is that this behavior is primarily or wholly attributable to the natural internal variability of the climate system. Other interpretations of the “hiatus” postulate that the relatively muted surface and tropospheric warming over the past 18 years is not due to a single cause but instead arises from the combined effects of internal variability and external forcing.

The external factor that has received a lot of attention in “hiatus” studies is the 21st century increase in volcanic aerosol forcing.13,14 Recently, Santer et al.15 detected signals of 21st century volcanic forcing in observations of the temperature of the lower troposphere and net clear-sky short-wave radiation at the top of the atmosphere. We considered whether volcanic signal detection is feasible for four additional observational variables: sea surface temperature, the temperature of the middle to upper troposphere, column-integrated water vapor, and precipitation.

Scientific Approach and Accomplishments

We use the climate simulations from the repository of CMIP5 that was the basis for the fifth Assessment Report of the Intergovernmental Panel on Climate Change, which contains simulations from different groups that are initialized every five years from 1960 to 2009. For a global climate model and a climate variable, Figure 1 shows observational values in black, uninitialized (long-term) global climate model simulation in blue and white, and different initializations of the decadal (shorter term) simulations in red.


Figure 2. ten-year prediction from 1989 to 1997 using observations and decadal simulations. the solid black line shows the observations used in the analysis and the dotted black line is the set of observations left out for prediction validation. the solid green line is the posterior mean estimate, the darker green shade is one standard deviation from the mean, and the lighter green is two standard deviations away from the mean.
Figure 1. Shown here are the monthly near surface temperature anomalies for observations (black) and the global climate model CanCM4, with the uninitialized mean (blue), individual realizations (white); and decadal realizations (red) for the initialization years 1979 to 2009. The decadal realizations are initialized every five years.
 

This kind of data can be used to gauge the variability within and between different climate models, as well as to assess the simulations’ ability to reproduce the historical record. We developed a coherent framework for evaluating multi-model decadal predictions and their blending into a unified probabilistic forecast. There is no fully established strategy for the weighting of the projections from multiple models based on their historical performance.4 Reliability Ensemble Average16 is a popular approach to assess simulation ensembles using observational records. Reliability Ensemble Average consists of calculating a weighted average of the ensemble members, where the weights are the result of a recursive calculation that uses a measure of the simulation’s ability to reproduce the observations. Bayesian hierarchical models offer a formal justification for the calculation of the weights.17 They also provide a coherent framework to account for all the relevant sources of variability and provide a probabilistic measure of the predictive uncertainty. The Bayesian hierarchical approach has been developed in a number of papers.18,19, 20, 21 Building on this tradition, we consider a hierarchical model that starts with an observation equation that postulates that the historical records provide noisy information about the climate. Likewise, the different climate simulations capture the climate signal, but are subject to discrepancies due to simulation inaccuracies. A summary of the information about the climate provided by the different sources of data is given by the posterior distribution, which serves to quantify the uncertainty of all the unobservable quantities in the statistical model. We developed a statistical model to blend/fuse different sources of information and produce probabilistic multi-model decadal climate predictions that reflect the accuracy of each member in how it reproduces the past decadal climate patterns.

We assume that the underlying climate process is common throughout all sources of information, with the different climate models having an additional model structure error or model bias component that is unique to the climate model. Figure 2 shows ten-year predictions from 1989 to 1997 using observations and decadal simulations. Similarly, by postulating a Bayesian hierarchical model we can obtain probabilistic predictions for other climate variables.22


Figure 1. shown here are the monthly near surface temperature anomalies for observations (black) and the global climate model cancm4, with the uninitialized mean (blue), individual realizations (white); and decadal realizations (red) for the initialization years 1979 to 2009. the decadal realizations are initialized every five years.
Figure 2. Ten-year prediction from 1989 to 1997 using observations and decadal simulations. The solid black line shows the observations used in the analysis and the dotted black line is the set of observations left out for prediction validation. The solid green line is the posterior mean estimate, the darker green shade is one standard deviation from the mean, and the lighter green is two standard deviations away from the mean.
 

We found that one contributing factor to the “warming hiatus” is an increase in volcanically induced cooling over the early 21st century.15 Volcanic signals are statistically detectable in multiple observed climate variables: in spatial averages of tropical and near-global sea surface temperature, tropospheric temperature, net clear-sky short-wave radiation, and atmospheric water vapor. Signals of late 20th and early 21st century volcanic eruptions are also detectable in near-global averages of rainfall. In tropical average rainfall, however, only a Pinatubo-caused drying signal is identifiable. Successful volcanic signal detection is critically dependent on removing variability induced by the El Nino–Southern Oscillation.

We used updated and improved satellite retrievals of the temperature of the mid- to upper troposphere to address key questions about the size and significance of trends in this part of the atmosphere, agreement with model-derived values, and whether models and satellite data show similar vertical profiles of warming.23 The average ratio of modeled and observed middle to upper troposphere temperature trends is sensitive to both satellite data uncertainties and to model-data differences in stratospheric cooling.

Impact on Mission

The growing societal interest in climate change has generated a need to understand near-term climate predictions that can assist in the decision-making that affects time scales of about a decade. Data are fundamental to all of science. Data enhance scientific theories, and their statistical analysis suggests new avenues of research and data collection. Providing regional-scale climate predictions on a timescale of several decades that are suitable for general climate-change risk assessment is well aligned with the Laboratory’s strategic focus area in energy and climate to understand climate challenges and develop options for future adaption, including decadal climate projections at the regional scale. Mining data from climate simulations also aligns well with the core competency in information systems and data science.

Through this LDRD we have begun developing a relationship between climate scientists and statisticians at Livermore. We have also brought in two new hires and have established collaborations with the University of California, Santa Cruz (UCSC). One of the new hires, Francisco Beltrán, started as a postdoc on the LDRD and has successfully transitioned to applying statistical methods for analysis of data that supports the National Ignition Facility. The other hire, Giuliana Pallotta, is continuing to work on data science problems in the Program for Climate Model Diagnosis and Intercomparison and is also broadening her research area to include National Atmospheric Release Advisory Center applications. Further, we have established a collaboration with UCSC Professor Bruno Sansó, who participated in Livermore’s visiting scientist program. Two of his students spent two summers in our summer intern program, getting exposure not only to data science problems in climate, but also to broader application areas.

Conclusion

This LDRD project has only scratched the surface of climate science problems to which statisticians could contribute. To quote from the paper "American Statistical Association’s Advisory Committee for Climate Change Policy" coauthored by our collaborator Professor Bruno Sansó: “Earth’s climate system is complex, involving the interaction of many different kinds of physical processes and many different time scales. Thus this area of science has a critical dependence on the examination of all relevant data and the application of statistics for its interpretation. Climate datasets are increasing in number, size, and complexity and keep challenging the traditional methods of data analysis. Most standard software is overwhelmed by the complexity of the data and of the physical processes that drive climate models, leaving climate scientists without the benefit of powerful statistical inferences. Further, identifying the signal and filling in gaps in these noisy, massive, yet incomplete datasets is a problem that should be tackled by both climate scientists and statistical scientists in close collaboration.”

Developing new statistical approaches is an essential part of understanding climate and its impact on society in the presence of uncertainty. Crucial to success are new statistical methods that recognize uncertainties in the measurements and the scientific processes but are also tailored to the unique scientific questions being studied.

The Laboratory needs to establish an interdisciplinary research program around climate, where statisticians have the opportunity to collaborate with researchers from other disciplines and advance the understanding of the climate system (e.g., quantification of uncertainties, the development of powerful tests of scientific hypotheses).

The Laboratory also needs to involve scientists and statisticians in partnerships or in teams to address problems in climate science. This program should also address the critical need for training a new generation of interdisciplinary researchers who can address challenging scientific problems that require complex data analysis by developing and using the necessary sophisticated statistical methods.

References

  1. Taylor, K. E., R. J. Stouffer, and G. A. Meehl, “An overview of CMIP5 and the experiment design.” Amer. Meteor. Soc. 93, 485 (2012).
  2. Meehl, G. A. “Decadal prediction.” Amer. Meteor. Soc. 90, 1467 (2009).
  3. Murphy J., et al., “Towards prediction of decadal climate variability and change.” Procedia Env. Sci. 1, 287 (2010).
  4. Bindoff, N. L., et al., “Detection and attribution of climate change: From global to regional,” Climate change 2013: The physical science basis. Cambridge University Press, Cambridge, United Kingdom (2013).
  5. Christy, J. R., Testimony in hearing before the U.S. Senate Committee on Commerce, Science, and Transportation, Subcommittee on Space, Science, and Competitiveness, Dec. 8, 2015.
  6. Mears, C., and F. J. Wentz, “Sensitivity of satellite-derived tropospheric temperature trends to the diurnal cycle adjustment.” Clim. 29, 3629 (2016).
  7. Zou, C.-Z., et al., “Recalibration of microwave sounding unit for climate studies using simultaneous nadir overpasses.” Geophys. Res. 111 (2006).
  8. Santer, B. D., et al, "Detection of climate change and attribution of causes," Contribution of working group I to the second assessment report of the Intergovernmental Panel on Climate Change. Cambridge University Press, Cambridge, United Kingdom, p. 407 (1995).
  9. Santer, B. D., et al., “Interpreting differential temperature trends at the surface and in the lower troposphere.” Science 287, 1227 (2000).
  10. Fu, Q., S. Manabe, and C. M. Johanson, “On the warming in the tropical upper troposphere: Models versus observations.” Res. Lett. 38 (2011).
  11. Karl, T. R., et al., Temperature trends in the lower atmosphere: Steps for understanding and reconciling differences. A Report by the U.S. Climate Change Science Program and the Subcommittee on Global Change Research. National Oceanic and Atmospheric Administration, Silver Springs, MD (2006).
  12. Santer, B. D., et al., “Separating signal and noise in atmospheric temperature changes: The importance of timescale.” Geophys. Res. 116 (2011).
  13. Solomon, S., et al., “The persistently variable ‘background’ stratospheric aerosol layer and global climate change.” Science 333, 866 (2011).
  14. Vernier, J.-P., et al., “Major influence of tropical volcanic eruptions on the stratospheric aerosol layer during the last decade.” Res. Lett. 38 (2011).
  15. Santer, B. D., et al., “Observed multivariable signals of late 20th and Early 21st century volcanic activity.” Res. Lett. 42(2), 500 (2015).
  16. Giorgi, F., and L. O. Mearns, “Calculation of average, uncertainty range, and reliability of regional climate changes from AOGCM simulations via the Reliability Ensemble Averaging (REA) method.” J. Clim. 15, 1141 (2002).
  17. Tebaldi, C., and Sansó, B., “Joint projections of temperature and precipitation change from multiple climate models: A hierarchical Bayesian approach.” Royal Stat. Soc. A 172, 83 (2009).
  18. Beltrán, F., et al., “Joint projections of north Pacific sea surface temperature from different global climate models.” Environmetrics 23, 451 (2012).
  19. Chandler, R. E., “Exploiting strength, discounting weakness: combining information from multiple climate simulators.” Trans. Royal Soc. London A 371 (2013).
  20. Harris, G. R., et al., “Probabilistic projections of transient climate change.” Clim. Dynam. 40, 2937 (2013).
  21. Sexton, D. M. H., et al., “Multivariate probabilistic projections using imperfect climate models part I: Outline of methodology.” Clim. Dynam. 38, 2513 (2011).
  22. Beltrán, F., Joint probabilistic projections of climate model simulations on decadal time scales. (2016). LLNL-TR-705860.
  23. Santer, B. D., et al., “Comparing tropospheric warming In climate models and satellite data.” J. Clim. 30(1), 373 (2017). http://dx.doi.org/10.1175/JCLI-D-16-0333.1

Publications and Presentations

  • Beltran, F. M., B. D. Santer, and A. Kupresanin, Decadal climate prediction using Bayesian Methods. Joint Statistical Mtgs. Chicago, IL, Jul. 30–Aug. 4, 2016. LLNL-POST-671559.
  • Beltran, F. M., B. D. Santer, and G. Johannesson, Effects of volcanic aerosols on lower troposphere temperature. Intl. Society for Bayesian Analysis, Sardinia, Italy, June 13–17, 2016. LLNL-PRES-676046.
  • Beltran, F. M., et al., Measuring the effects of volcanic aerosols on different climate variables. (2016). LLNL-TR-706822.
  • Beltran, F. M., B. D. Santer, and G. Johannesson, Quantifying the impact of volcanic aerosol forcing uncertainties on lower troposphere temperature. Joint Statistical Mtg., Seattle, WA, Aug. 8–13, 2015. LLNL-ABS-799361.
  • Sanso, B., et al., Assessment and blending of decadal climate simulations. Intl. Society for Bayesian Analysis, Sardinia, Italy, June 13–17, 2016. LLNL-PRES-694783.
  • Santer, B. D., et al., “Volcanic effects on climate.” Nat. Clim. Change 6, 3 (2016). LLNL-JRNL-673787. http://dx.doi.org/doi:10.1038/nclimate2859
  • Santer, B. D., et al., “Observed multivariable signals of late 20th and early 21st century volcanic activity.” Geophys. Res. Lett. 2, 500 (2015). LLNL-JRNL-660338. http://dx.doi.org/10.1002/2014GL062366
  • Santer, B. D., et al., Climate impact of volcanic forcing uncertainty. (2015). LLNL-JRNL-677157.
  • Santer, B. D., et al., “Volcanic effects on climate.” Nat. Clim. Change 6, 3 (2015). LLNL-JRNL-673787.