z-logo
Premium
Model averaging to combine simulations of future global vegetation carbon stocks
Author(s) -
Butler Adam,
Doherty Ruth M.,
Marion Glenn
Publication year - 2009
Publication title -
environmetrics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.68
H-Index - 58
eISSN - 1099-095X
pISSN - 1180-4009
DOI - 10.1002/env.953
Subject(s) - environmental science , gcm transcription factors , probabilistic logic , climatology , vegetation (pathology) , climate model , general circulation model , climate change , regression , range (aeronautics) , bayesian probability , baseline (sea) , statistical model , meteorology , atmospheric sciences , statistics , mathematics , geology , geography , medicine , oceanography , materials science , pathology , composite material
We quantify the impact of climate model uncertainty upon predictions of future vegetation carbon stocks, for the period up to 2100, generated by a dynamic global vegetation model (DGVM) under a particular emissions scenario (SRES A2). Deterministic simulations are generated from the Lund‐Potsdam‐Jena (LPJ) model using climatic inputs derived from nine general circulation models (GCMs). Simulated changes between 1961–1990 and 2070–2099 range from +26 to +133 gtC, and are in broadly good agreement with those obtained in a recent previous study using the LPJ model. Simulated values for the 20th century are also obtained by running LPJ with observed climate data, and this provides a baseline against which the other runs can be compared. Time series regression models are used to analyse the discrepancies between each of the GCM‐based simulations and the baseline simulation, and a novel form of model averaging—in which we average not only across GCM‐based simulations but also across models for each discrepancy—is then used to combine these into a single probabilistic projection for global stocks of vegetation carbon. Weights for the regression models are estimated in a simple post hoc way using the Bayesian information criterion (BIC), and the weights for the GCMs are either estimated in the same way or else fixed to be equal. Estimating the GCM weights leads the predictions to be dominated by a single climate model and hence produces narrow predictive distributions. If GCMs are weighted equally then the predictive distributions are much more diffuse and span the full range of simulated values. Copyright © 2008 John Wiley & Sons, Ltd.

This content is not available in your region!

Continue researching here.

Having issues? You can contact us here