Improving efficiency and effectiveness of Bayesian recursive parameter estimation for hydrologic models

Abstract

There are several sources of uncertainties in hydrologic modeling studies. Conventional deterministic modeling techniques typically ignore most of these uncertainties. However, there has been a growing need for better quantification of the accuracy and precision of hydrologic model predictions. Bayesian Recursive Estimation (BaRE) is an algorithm being developed towards considering these uncertainties for parameter estimation and prediction within an operational setting. This dissertation work evaluated and improved the current version of the algorithm. The methodology was improved using a progressive re-sampling of the Highest Probability Density (HPD) region of the parameter space, which concentrated the samples in the current HPD region while terminating computations in the nonproductive portions of the parameter space, rather than evaluating feasible parameter space based on the initial set of samples. The covariance structure of the well behaving parameter sets is used to generate new parameter sets, resulting in significant improvements compared to the original BaRE. Further, to reduce the "model/data overconfidence" problem, an entropy term and a data lack-of-confidence factor were introduced into the probability-updating rule. Comparison to batch calibration using the popular Shuffled Complex Evolution (SCE-UA) optimization method indicated that the improved recursive calibration technique is a powerful tool, especially useful where basins are recently gauged and hydrologic data are not well accumulated. The final method is also effective in tracing the temporal variations of parameters as a response to natural or human induced changes in the hydrologic system.