Parameter Sensitivity Analysis for Computationally Intensive Spatially Distributed Dynamical Environmental Systems Models

Dynamical environmental systems models are highly parameterized, having large numbers of parameters whose values are uncertain. For spatially distributed continental-scale applications, such models must be run for very large numbers of grid locations. To calibrate such models, it is useful to be able to perform parameter screening, via sensitivity analysis, to identify the most important parameters. However, since this typically requires the models to be run for a large number of sampled parameter combinations, the computational burden can be huge. To make such an investigation computationally feasible, we propose a novel approach to combining spatial sampling with parameter sampling and test it for the Noah-MP land surface model applied across the continental United States, focusing on gross primary production and flux of latent heat simulations for two vegetation types. Our approach uses (a) progressive Latin hypercube sampling to sample at four grid levels and four parameter levels, (b) a recently developed grouping-based sensitivity analysis approach that ranks parameters by importance group rather than individually, and (c) a measure of robustness to grid and parameter sampling variability. The results show that a relatively small grid sample size (i.e., 5% of the total grids) and small parameter sample size (i.e., 5 times the number of parameters) are sufficient to identify the most important parameters, with very high robustness to grid sampling variability and a medium level of robustness to parameter sampling variability. The results ensure a dramatic reduction in computational costs for such studies.