An efficient method for global parameter sensitivity analysis and its applications to the Australian community land surface model (CABLE)

Abstract

State-of-the-art global land surface models (LSMs) have a large number (i.e. a few hundred) of parameters. Many of those parameters are poorly constrained and are therefore very uncertain. Usually only a few of the parameters are responsible for changes in the model output of interest. Identifying those parameters that have a significant effect on the model output is an important step before applying parameter estimation methods using observations. However this has not been done systematically for any global LSMs yet, because of the computational costs involved. Here, we introduce a global sensitivity analysis method that is widely used in chemical engineering. This method includes two steps: a screening step that ranks all model parameters by their importance on model output in order to select the potentially important parameters and a second step that aims to quantify the contribution to the variance of model output by each of the pre-selected parameters and by their interactions. This method can be readily applied to any model. Here we apply this method to the Australian community land surface model (CABLE) as an example, and find that the two-step approach is efficient as only 690 model simulations are required to identify the few important parameters amongst the 22 parameters for each of the 10 plant functional types (PFTs) in the first step. Another 256 model simulations are required for the variance based analysis in the second step. We find that the leaf maximum carboxylation rate (vcmax) is by far the most important parameter for global annual gross primary productivity (GPP) across all PFTs. However, if focusing on annual latent heat flux (LE) the importance of the parameters is very much PFT dependent. We suggest that this two-step approach should be used to identify important parameters in global LSMs, so that observations to constrain parameters can be used more efficiently in a subsequent parameter estimation step.