Quantifying the Sensitivity of Energy Fluxes to Land Surface Parameter Selection Using the Active Subspace Method

Abstract:

Complex hydrologic models are commonly used as computational tools to assess and quantify fluxes at the land surface and for forecasting and prediction purposes. When estimating water and energy fluxes from vegetated surfaces, the equations solved within these models require that multiple input parameters be specified. Some parameters characterize land cover properties while others are constants used to model physical processes like transpiration. As a result, it becomes important to understand the sensitivity of output flux estimates to uncertain input parameters. The active subspace method identifies the most important direction in the high-dimensional space of model inputs. Perturbations of input parameters in this direction influence output quantities more, on average, than perturbations in other directions. The components of the vector defining this direction quantify the sensitivity of the model output to the corresponding inputs. Discovering whether or not an active subspace exists is computationally efficient compared to several other sensitivity analysis methods. Here, we apply this method to evaluate the sensitivity of latent, sensible and ground heat fluxes from the ParFlow-Common Land Model (PF-CLM). Of the 19 input parameters used to specify properties of a grass covered surface, between three and six parameters are identified as important for heat flux estimates. Furthermore, the 19-dimenision input parameter space is reduced to one active variable and the relationship between the inputs and output fluxes for this case is described by a quadratic polynomial. The input parameter weights and the input-output relationship provide a powerful combination of information that can be used to understand land surface dynamics. Given the success of this proof-of-concept example, extension of this method to identify important parameters within the transpiration computation will be explored.