Application of Three Parameter Estimation Methods of Hydrologic Model: Uncertainty in the Parameter Calibration

摘要

Parameter calibration is a critical step in hydrologic model applications. Many parameter calibration techniques have been developed during the past two decades. Among them, heuristic optimization algorithms and Bayesian inference theory based parameter estimation methods are widely used. The former approach provides point estimations of parameters of hydrologic models and the latter gives interval estimations of the parameters. The Bayesian approach becomes more active in recent years because of the uncertainty estimations it provides. How does the Bayesian approach compare with the heuristic optimization approach in estimating hydrologic model parameters? To investigate this problem, three parameter calibration procedures, SCE-PSO, GLUE and SCEM-UA, are employed to estimate parameters associated with a hydrologically based land surface model, VIC-3L. SCE-PSO is a newly developed heuristic optimization algorithm which couples the strengths of the Shuffled Complex Evolution algorithm (SCE-UA) developed by Duan (1992) and the Particle Swarm Optimization (PSO) developed by Eberhat and Kennedy in 1990s. SCE-PSO has good performance in global search and it can provide point estimation for model parameters through calibration. The Generalized likelihood uncertainty estimation (GLUE) methodology for hydrologic models is based on the Bayesian inference framework and can provide interval estimation, sensitivity analysis and uncertainty assessment. SCEM-UA, developed by Vrugt (2003), is a Markov Chain Monte Carlo (MCMC) sampler which integrates features of SCE-UA, Bayesian statistics, and Metropolis algorithm. As a MCMC sampler, the sampled feasible space needs to be ergodic. The three parameter optimization procedures were applied to a watershed to estimate the model parameters in VIC-3L. The forcing data of VIC-3L and stream flow data used in calibration are obtained from the MOPEX project. The issues of equifmality and parameter interactions, which are important sources of uncertainty in hydrologic model parameter estimation, were investigated and major findings are presented as well.