ABSTRACT OF THE THESIS:

Information Content of Hydrologic Signatures and their Impacts on Watershed Model Calibration

By

Mohammad Sassani

Master of Science in Environmental Engineering

University of California, Irvine, 2018

Professor Amir AghaKouchak, Chair

Approximate Bayesian computation, also known as ABC, is a simulation-based method with its roots in Bayesian analysis, which is increasingly used in the past couple of years in the field of hydrology for parameter estimation and uncertainty analysis. ABC relaxes the need for an explicit likelihood function and uses multiple hydrologic metrics (signatures) to evaluate model simulations. If the hydrologic signatures are sufficient, and the distance between simulated metrics and their observed counterparts are less than a nominal threshold, model parameters are assumed to be derived from the posterior distribution. Since finding a sufficient set of summary statistics is difficult for complex environmental systems and high-order models, the inferred posterior parameter and predictive distributions are constrained on the set of selected metrics and may not necessarily mirror the "true" posterior parameter distributions.

In this thesis, we have strived to shed some light on how different hydrologic signatures constrain watershed models. More specifically, we are interested to understand which portions of the hydrograph is constrained by each individual hydrologic signature, and ultimately move toward finding a set of sufficient metrics with least overlapping information. Throughout the years significant strides have been made in the field of calibration and uncertainty assessment of hydrologic models, but the constraining power of different hydrologic metrics have not yet been characterized. Using information theory, likelihood, entropy, and conventional metrics such as NSE, RMSE, and percent bias; we try to characterize constraining power and model parameter sensitivity to each metric. Sensitivity analysis of model parameters is performed using Kullback-Leiber divergence metric and Two Sample Kolmogorov Smirnov test. One interesting finding of this study is that hydrologic signatures are model specific and cannot be readily transferred between different models. The result of this study also depict that that each watershed possesses different and unique characteristics, which is also manifested in the hydrologic signatures that are selected for each watershed. In other word, hydrologic signatures are model- and watershed- specific. We also observe that more flexible watershed models, such as GR4J, react differently to hydrologic signatures as compared to less flexible model structures, such as HyMod. In more detail, more flexible model structures are associated with higher information content, which might ultimately help the hydrologic signatures in constraining model behavior.