Developing climate model ensembles: A comparative case study

Abstract

Projections of future changes in hydroclimatic variables are available through various general circulation models (GCMs) that are part of the Coupled Model Intercomparison Project Phase 5 (CMIP5). Assessment of the performance of these models in simulating past climate, both individually and as an ensemble, has received much attention and is commonly the first step in assessing their suitability of application. In this study, we have evaluated the ability of various ensemble models to simulate past temperature and precipitation in the Gulf Basin region of North America, chosen as an illustrative case study. We have developed ensembles from 34 CMIP5 GCMs using six diverse approaches, including random forest, support vector regression, neural networks, linear regression, and weighted k-nearest neighbors, and compared the performance of the ensembles with each other and the individual GCMs using a robust set of metrics and nonparametric tests. For temperature, random forest outperforms all other ensembles and the best-performing GCM by a statistically significant margin and is able to simulate temporal and spatial patterns in temperature well. None of the ensembles are able to adequately simulate observed precipitation patterns in the study area, likely due to spatial differences in precipitation drivers in the region, as well as the coarseness of the dataset itself. However, the random forest, support vector regression, neural network, and linear regression ensembles achieved statistically significant improvements to precipitation simulation, as compared to the individual GCMs and the simple arithmetic mean ensemble, which has been used in several studies.