Streamflow, stomata, and soil pits: sources of inference for complex models with fast, robust uncertainty quantification


The scale and complexity of environmental and earth systems introduce an array of uncertainties that need to be systematically addressed. In numerical modeling, the ever-increasing complexity of representation of these systems confounds our ability to resolve relevant uncertainties. Specifically, the numerical representation of the governing processes involve many inputs and parameters that have been traditionally treated as deterministic. Considering them as uncertain introduces a large computational burden, stemming from the requirement of a prohibitive number of model simulations. Furthermore, within hydrology, most catchments are sparsely monitored, and there are limited, heterogeneous types of data available to confirm the model’s behavior. Here we present a blueprint of a general approach to uncertainty quantification for complex hydrologic models, taking advantage of recent methodological developments. We rely on polynomial chaos machinery to construct accurate surrogates that can be efficiently sampled for the ecohydrologic model tRIBS-VEGGIE to mimic its behavior with respect to a selected set of quantities of interest. The use of the Bayesian compressive sensing technique allows for fewer evaluations of the computationally expensive tRIBS-VEGGIE. The approach enables inference of model parameters using a set of observed hydrologic quantities including stream discharge, water table depth, evapotranspiration, and soil moisture from the Asu experimental catchment near Manaus, Brazil. The results demonstrate the flexibility of the framework for hydrologic inference in watersheds with sparse, irregular observations of varying accuracy. Significant computational savings imply that problems of greater computational complexity and dimension can be addressed using accurate, computationally cheap surrogates for complex hydrologic models. This will ultimately yield probabilistic representation of model behavior, robust parameter inference, and sensitivity analysis without the need for greater investment in computational resources.

Advances in Water Resources