CSDMS 2014 annual meeting poster Xuan Yu

Presentation provided during CSDMS annual meeting 2014

Parameter and model uncertainty analysis of a physics-based hydrologic model: a comparative study of GLUE and Gaussian process emulator

Xuan Yu, Pennsylvania State University University Park Pennsylvania, United States. xuanyupsu@gmail.com
Christopher Duffy, Pennsylvania State University University Park Pennsylvania, United States. cxd11@psu.edu
Murali Haran, Pennsylvania State University University Park Pennsylvania, United States. muh10@psu.edu

CSDMS2014 poster XuanYu.png


Quantification of uncertainty of environmental models plays an important role in the decision making process. The most popular method GLUE (Generalized Likelihood Uncertainty Estimation) has been argued for the computationally inefficiency due to the emerging of complex models. Recently, the Bayesian approach using Gaussian process (GP) emulator has been attracted much attention in the uncertainty analysis of computationally expensive models. It would be useful to compare the difference of this two methods in the uncertainty analysis of a physics-based hydrologic model. We evaluate the difference of GLUE and GP emulator in the assessment of parameter uncertainty of a physics-based integrated hydrologic model (Penn State Integrated Model: PIHM). PIHM integrates the hydrological processes including interception, throughfall, infiltration, recharge, evapotranspiration, overland flow, groundwater flow, and channel routing, in a fully coupled scheme. The tradition parameter estimation focuses only on the model performance at streamflow, which may cause significant uncertainty of parameter in other intermediate predicted variables (groundwater table, soil moisture, etc.). We demonstrate the uncertainty at each process to investigate the uncertainty transfer in the integrated framework of PIHM. This study considers the comparison between GLUE and GP emulator uncertainty analysis at a catchment at central PA: Shale Hills, where hydrologic processes are monitored, including intermediate variables of rainfall-runoff processes. The GLUE method attempts to evaluate the total uncertainty from model structure, parameter, and input data through a large sample of model simulation. Whereas, the GP emulator starts with designed samples of model runs, and allows inferring the values of the model output at untried points. The uncertainty disentanglement between model structure, parameter, and hydrologic processes suggests that ignoring intermediate variable uncertainty will lead to unrealistic model simulations.

* Please acknowledge the original contributors when you are using this material. If there are any copyright issues, please let us know and we will respond as soon as possible.