Simulating the Effect of Uncertain Model Drivers on Hydrologic Predictions via an Approximate Bayesian Approach

Abstract:

Understanding the extent and effect of observational uncertainty remains a key concern in hydrologic model specification. Advances have been made in developing robust Bayesian approaches for characterizing the impact of uncertain climatological drivers on hydrologic predictions and parameters. However, these approaches are typically very high dimensional, requiring specification of large numbers of variables that represent statistical uncertainty in the model inputs. Recent developments in approximate Bayesian methods offer an elegant alternative to the fully Bayesian approach. Approximate Bayesian Computation (ABC) is commonly used in situations where a model is easy to simulate from, but where the likelihood is difficult or impossible to calculate. The ABC approach provides an opportunity to develop novel hydrologic statistics for model inference and to develop efficient methods for parameter identification in high dimensional hydrologic models. In this study, we demonstrate the use of approximate Bayesian methods for characterizing uncertain model inputs across multiple hydrologic case studies. Model inference is conducted via statistics that capture hydroclimatic and hydrologic functioning. Our analysis investigates the utility of ABC for model assessment, parameter identification and uncertainty characterization when dealing with potentially large observational uncertainties in hydroclimatic applications.