Improving Multi-model Ensembles of Climate Projections through Time Variability Correction and Ensemble Dependence Transformation

The uncertainty of future climate change is typically estimated by a collection of models from various climate institutions. The Ensemble Dependence Transformation (EDT) method has proven effective in producing ensembles with means that lie closer to the verifying observations and with variances that match the variance of the observations about the ensemble mean. However, EDT does not specifically address temporal variability and persistence attributes within individual models. This limitation can potentially be addressed by the Time Variability Correction (TVC) method, designed to quantify, and correct model variability errors across differing time scales.

In this study, we test and compare four approaches: 1) TVC only; 2) EDT only; 3) TE: Applying TVC to individual models first, followed by EDT; 4) ETE: Applying EDT to obtain the time-varying mean series, using TVC to correct the EDT-transformed series, and applying EDT again to TVC post-processed model series. These methods are employed to post-process 26 CMIP6 (Coupled Model Intercomparison Project Phase 6) daily mean temperature projections across Australia under a model-as-truth setup.

We evaluate the results using verification metrics for assessing both individual models and multi-model ensembles. Findings indicate that, overall, ETE performs better in improving individual model statistics, including variance and lag correlations relative to the time-varying ensemble mean. Additionally, ETE enhances ensemble statistics, notably ensemble standard deviation (ESD) during both in-sample historical and out-of-sample projection periods. TE is particularly effective at improving root mean squared difference (RMSD) between ensemble mean and observations, along with continuous ranked probability skill score (CRPSS).