Simulating Hydrologic Flux Sensitivity to Model Parameters in the Community Land Model
The integrated Earth system involves multi-phase, multi-component biogeophysical and biogeochemical processes and integrated models introducing numerous model and coupling parameters and therefore a formidable high-dimensional parameter space. Because many of the parameters cannot be observed or measured, they are subject to great uncertainty, making it difficult to predict and quantify uncertainty and risk for decision-making. Uncertainty quantification (UQ) and sensitivity analysis (SA) approaches are used to narrow the range of uncertainty, yet their use may yield completely different conclusions toward understanding of the complex system.
Various widely-accepted SA methods were systematically evaluated and compared; meanwhile, the effects of linear, interaction, and high-order terms of input variables were fully explored and narrowed down to fewer terms using parameter reduction/selection techniques. This study provides guidance on the optimal SA approaches as well as on the strategy for optimizing the Community Land Model parameters to improve predictions of land surface fluxes.
To evaluate how various metrics and/or SA methods might affect the SA results, a research team led by Department of Energy scientists at Pacific Northwest National Laboratory explored four different approaches for hydrologic fluxes at MOPEX 07147800 watershed, located at Winfield, Walnut River, Kansas. Generally, parameters (such as specific yield) were identified to dominate deviations between observed hydrological fluxes and simulations from the Version 4 of the Community Land Model (CLM4). Also, adopting different SA approaches was shown to yield different parameter optimization strategies, and such differences were more obvious for parameters whose contributions are weaker than the dominant ones. The common parameters identified as important by different SA approaches are of high priority to be optimized for reducing the discrepancies between runoff/latent heat observations and model simulations. The research recommends conducting model optimization using multiple data or metrics with supplementary information to each other.