Large uncertainties remain in climate predictions, many of which originate from uncertainties in land-surface processes. In particular, uncertainties in land-atmosphere fluxes of carbon dioxide and energy are driven by incomplete knowledge about model parameters and their variation over space and time. Using the ACME land model, we perform uncertainty decomposition based on global Polynomial Chaos (PC) surrogate construction, using the Bayesian Compressive Sensing (BCS) sparse learning technique.