Probabilistic projection of regional climate change based on CMIP5 long-term experiments

Climate models have been widely used to study the anthropogenic influence on the climate and to project the future changes under certain emission scenarios. Among the approaches to combine multi-model ensemble projection, the Bayesian framework has shown its advantage as it takes account of the relationship between observations and model simulations. Moreover, a probabilistic projection can be derived from the Bayesian posterior distribution. Outputs from models participating in the fifth phase of the Coupled Model Intercomparison Project (CMIP5) are becoming available recently, and we use them to project the regional temperature change under new scenarios. A Bayesian statistical model was built from observations and outputs of historical and rcp85 experiments from 28 CMIP5 models. In this study, global land was divided into 21 regions and the seasonal mean temperature was aggregated over each region. A Gaussian distribution with a time-dependent location parameter was chosen for the likelihood component of the statistical model. The Bayesian posterior distribution was obtained through Markov Chain Monte Carlo method with a Gibbs sampler. The time-dependent location parameter allowed us to use longer time series of observation and simulations, thus made the chain easier to converge. To exclude model uncertainty in the projections, parameters indicating model bias were included. The prior distributions for all the parameters were uninformative except that for bias change. Through sensitivity experiments, we suggest that a wider prior distribution for bias change would cause the underestimate of the uncertainty. How to better restrain the bias change by incorporating other information like the uncertainty estimate for climate sensitivities will be discussed.