Uncertainty analysis of a semi-distributed hydrologic model based on a Gaussian Process emulator

Abstract

Despite various criticisms of GLUE (Generalized Likelihood Uncertainty Estimation), it is still a widely-used uncertainty analysis technique in hydrologic modelling that can give an appreciation of the level and sources of uncertainty. We introduce an augmented GLUE approach based on a Gaussian Process (GP) emulator, involving GP to conduct a Bayesian sensitivity analysis to narrow down the influential factor space, and then performing a standard GLUE uncertainty analysis. This approach is demonstrated for a SWAT (Soil and Water Assessment Tool) application in a watershed in China using a calibration and two validation periods. Results show: 1) the augmented approach led to the screening out of 14–18 unimportant factors, effectively narrowing factor space; 2) compared to the more standard GLUE, it substantially improved the sampling efficiency, and located the optimal factor region at lower computational cost. This approach can be used for other uncertainty analysis techniques in hydrologic and non-hydrologic models.