Characterizing Parameter Uncertainty of a Distributed Hydrologic Model via Basin Internal Observations

Andrea Thorstensen

NOAA/ESRL PSL

Tuesday, Mar 21, 2017, 2:00 pm
DSRC Room 2A-305

Abstract

Uncertainty characterization of hydrologic models becomes vital, particularly in the context of extreme events (i.e. floods). These extreme events are rare, which means the models constructed based on observed data will have very few samples of such events. This has consequences on model development as well as parameterization. To further complicate this problem, streamflow serves as the primary observation source for building and calibrating hydrologic models. The signals from streamflow are often modified by manmade structures such as dams, whose operators’ choices of reservoir release are not readily captured by the hydrologic model. As a consequence, model parameters that are often conceptual in nature may improperly compensate for release patterns during the calibration period. In response to this challenge, this study aims to leverage basin internal observations of streamflow and soil moisture in an effort to capture areas of modeling weakness for the Russian River Basin in Northern California in the context of a distributed hydrologic model designed for operational use. At the center of this study is the National Weather Service's Hydrology Laboratory Research Distributed Hydrologic Model (HL-RDHM) and its conceptually-based Sacramento Soil Moisture Accounting Model with Heat Transfer (SAC-HT) submodel for rainfall-runoff generation. The Russian River Basin is instrumented with 11 streamflow gauges from the US Geological Survey and 7 multi-depth NOAA Hydrometeorology Testbed soil moisture observation stations, with most records dating back to 2010. The multitude of streamflow gauges makes this basin of particular interest to characterize parameter uncertainty throughout the channel routing of the model, and the soil moisture observations allow for a check on model state behavior during sub-basin processes. Parameter calibration and uncertainty estimation is carried out via the DiffeRential Evolution Adaptive Metropolis (DREAM) algorithm, which is applied to the outlet streamflow, sub-basin streamflow, and in situ soil moisture observations. Results show that while soil water storage parameters calibrated using the outlet streamflow signal exhibit similar levels of uncertainty as those calibrated using the soil moisture signal, the optimal values and distributions of these two cases differ dramatically. This study serves as a cautionary tale of parameter uncertainty estimation in a distributed hydrologic model, given that different results from each scenario highlight the possibility of attributing uncertainty to the improper source.