Uncertainties in land surface processes contribute to a large spread in model predictions of climate change within the 21st century. Land surface models incorporate diverse processes across various temporal and spatial scales, and can include dozens of uncertain parameters. The signs and magnitudes of predicted fluxes are highly uncertain and depend on multiple feedbacks in terrestrial systems. Improved understanding about the sensitivity of model outputs to specific parameters and processes, and the contribution of parametric uncertainty to overall prediction uncertainty, is of critical importance for model development and improved predictions.

Global sensitivity analysis (GSA) is a useful tool for attributing prediction uncertainty to specific model parameters. Because these analyses require a large ensemble of model simulations, surrogate models, or simplified representations of complex models, have often been applied. Surrogates based on Polynomial Chaos (PC) expansions are convenient for GSA. However, the PC surrogate construction still requires a large ensemble of simulations, especially when the number of parameters is large. Here we apply a new procedure, the weighted Iterative Bayesian Compressive Sensing (WIBCS) algorithm, which allows a sparse, high-dimensional PC surrogate with very few model evaluations.

We use the WIBCS algorithm and GSA to determine the sensitivity the Accelerated Climate Model for Energy (ACME) Land Model to 65 parameters. Several model output variables are analyzed including gross primary productivity (GPP), leaf area index (LAI), vegetation carbon and soil organic matter carbon at nearly 100 FLUXNET sites covering a broad range of multiple plant functional types (PFTs) and climates. We find for all PFTs, generally 15 or fewer parameters drive most of the variance in the outputs. Within a PFT for a given output, generally the same parameters appear as sensitive at each site while differences in parameters are evident among PFTs and different outputs. The sensitivities of some parameters vary as a function of climate variables such as temperature or precipitation. This sensitivity analysis will serve as the basis for more focused, lower-dimensional studies leading to parameter calibration and improved land-surface model prediction at global scales.