Drift Correction and Sub-Ensemble Predictive Skill Evaluation of the Decadal Prediction Large Ensemble With Application to Regional Studies

Abstract

A large ensemble of experiments is required to reveal the predictable climate signal masked by the background noise in the decadal climate prediction (DCP). This is one of the main obstacles which complicates the generation of high-resolution decadal climate information at regional scale, given the computing cost of the task. In this study, a set of representative sub-ensembles of three members (ENS3) from the Decadal Prediction Large Ensemble has been selected to produce dynamically downscaled DCPs in future studies, minimizing the amount of computing resources required to conduct the regionalization while reducing as much as possible the loss of predictive skill with respect to the full ensemble (ENS40). The procedure to follow comprises two steps: first, an analysis to choose the most appropriate method of drift correction to remove the model drift; second, the selection of three members to build ENS3 and the evaluation of its performance and the impact of ensemble size on sub-ensemble performance. The study has been focused on sea surface temperature (SST), near-surface temperature anomaly and sea level pressure over some Coordinated Regional Climate Downscaling Experiment regions: Europe, South America and North America. The initial condition-based approach has been shown to be the most suitable method in the three domains. Although there is an inevitable loss of predictive skill when reducing the ensemble size, ENS3 has shown to be a relatively good alternative to ENS10 and ENS40 when facing computing constraints and the analysis is focused on SST.