HR: 09:15h
AN: H11G-05    [Abstracts]
TI: Bayesian analysis of data and model error in rainfall-runoff hydrological models
AU: * Kavetski, D
EM: kavetski@princeton.edu
AF: Princeton University, Dept of Civil and Environmental Engineering, Princeton, NJ 08544 United States
AU: Franks, S W
EM: stewart.franks@newcastle.edu.au
AF: University of Newcastle, School of Engineering, Callaghan, NSW 2308 Australia
AU: Kuczera, G
EM: george.kuczera@newcastle.edu.au
AF: University of Newcastle, School of Engineering, Callaghan, NSW 2308 Australia
AB: A major unresolved issue in the identification and use of conceptual hydrologic models is realistic description of uncertainty in the data and model structure. In particular, hydrologic parameters often cannot be measured directly and must be inferred (calibrated) from observed forcing/response data (typically, rainfall and runoff). However, rainfall varies significantly in space and time, yet is often estimated from sparse gauge networks. Recent work showed that current calibration methods (e.g., standard least squares, multi-objective calibration, generalized likelihood uncertainty estimation) ignore forcing uncertainty and assume that the rainfall is known exactly. Consequently, they can yield strongly biased and misleading parameter estimates. This deficiency confounds attempts to reliably test model hypotheses, to generalize results across catchments (the regionalization problem) and to quantify predictive uncertainty when the hydrologic model is extrapolated. This paper continues the development of a Bayesian total error analysis (BATEA) methodology for the calibration and identification of hydrologic models, which explicitly incorporates the uncertainty in both the forcing and response data, and allows systematic model comparison based on residual model errors and formal Bayesian hypothesis testing (e.g., using Bayes factors). BATEA is based on explicit stochastic models for both forcing and response uncertainty, whereas current techniques focus solely on response errors. Hence, unlike existing methods, the BATEA parameter equations directly reflect the modeler's confidence in all the data. We compare several approaches to approximating the parameter distributions: a) full Markov Chain Monte Carlo methods and b) simplified approaches based on linear approximations. Studies using synthetic and real data from the US and Australia show that BATEA systematically reduces the parameter bias, leads to more meaningful model fits and allows model comparison taking into account forcing uncertainty. The full MCMC approach also yields estimates of the true forcing (conditioned on the model assumptions), which can be used to improve data collection. We expect the ability to meaningfully disaggregate sources of uncertainty to be of significant benefit in hydrology and environmental modeling in general.
DE: 1854 Precipitation (3354)
DE: 1860 Runoff and streamflow
DE: 0644 Numerical methods
SC: Hydrology [H]
MN: 2004 AGU Fall Meeting