Climate Projections Using Bayesian Model Averaging and Space–Time Dependence

Abstract

Projections of future climatic changes are a key input to the design of climate change mitigation and adaptation strategies. Current climate change projections are deeply uncertain. This uncertainty stems from several factors, including parametric and structural uncertainties. One common approach to characterize and, if possible, reduce these uncertainties is to confront (calibrate in a broad sense) the models with historical observations. Here, we analyze the problem of combining multiple climate models using Bayesian Model Averaging (BMA) to derive future projections and quantify uncertainty estimates of spatiotemporally resolved temperature hindcasts and projections. One advantage of the BMA approach is that it allows the assessment of the predictive skill of a model using the training data, which can help identify the better models and discard poor models. Previous BMA approaches have broken important new ground, but often neglected space–time dependencies and/or imposed prohibitive computational demands. Here we improve on the current state-of-the-art by incorporating space–time dependence while using historical data to estimate model weights. We achieve computational efficiency using a kernel mixing approach for representing a space–time process. One key advantage of our new approach is that it enables us to incorporate multiple sources of uncertainty and biases, while remaining computationally tractable for large data sets. We introduce and apply our approach using BMA to an ensemble of Global Circulation Model output from the Intergovernmental Panel on Climate Change Fourth Assessment Report of surface temperature on a grid of space–time locations.