Keywords
runoff estimation, regionalisation, ungauged catchments, model averaging, simhyd
Start Date
1-7-2006 12:00 AM
Abstract
The problem of estimating runoff in ungauged catchments remains an important but elusive one. Previous studies suggest that there are two important properties common to rainfall-runoff models: over-parameterisation, leading to parameter covariance; and equifinality, the existence of multiple parameter sets which reproduce the streamflow adequately. Both reduce parameter identifiability, impeding identification of relationships between model parameter values and catchment characteristics that would otherwise be useful for regionalisation. This study investigates the use of a model averaging framework to circumvent this problem. Multiple behavioural parameter sets arising from Monte-Carlo simulation are used from each catchment in order to retain information about data and parameter uncertainty, and to estimate the uncertainty of prediction. The model averaging is based simply on each catchment’s physical similarity to a target ungauged catchment. Ungauged prediction results are assessed based on the Nash-Sutcliffe Efficiency (E). The model averaging schemes are compared to local cross-validation results (the benchmark), ‘nearest neighbour’ calibration (parameter sets taken from calibration of the geographically closest gauged catchment) and ‘regression method’ (each parameter value is estimated using regressions between optimised parameter values in gauged catchments and catchment characteristics). The study is carried out using the conceptual daily rainfall-runoff model SIMHYD on 44 catchments in south-east Australia. The results indicate that the model averaging approach shows promise for estimating streamflow in ungauged catchments. The streamflow simulations are significantly better when model parameter sets are retained using the model averaging approach than when parameter values are estimated using the nearest neighbour and regression approaches. The results from the model averaging approach are also better than the cross-validation results in over 40 % of the catchments. It is likely that more detailed analyses of the choice of weights and descriptors of catchment similarity will lead to even better modelling results.
Model Averaging, Equifinality and Uncertainty Estimation in the Modelling of Ungauged Catchments
The problem of estimating runoff in ungauged catchments remains an important but elusive one. Previous studies suggest that there are two important properties common to rainfall-runoff models: over-parameterisation, leading to parameter covariance; and equifinality, the existence of multiple parameter sets which reproduce the streamflow adequately. Both reduce parameter identifiability, impeding identification of relationships between model parameter values and catchment characteristics that would otherwise be useful for regionalisation. This study investigates the use of a model averaging framework to circumvent this problem. Multiple behavioural parameter sets arising from Monte-Carlo simulation are used from each catchment in order to retain information about data and parameter uncertainty, and to estimate the uncertainty of prediction. The model averaging is based simply on each catchment’s physical similarity to a target ungauged catchment. Ungauged prediction results are assessed based on the Nash-Sutcliffe Efficiency (E). The model averaging schemes are compared to local cross-validation results (the benchmark), ‘nearest neighbour’ calibration (parameter sets taken from calibration of the geographically closest gauged catchment) and ‘regression method’ (each parameter value is estimated using regressions between optimised parameter values in gauged catchments and catchment characteristics). The study is carried out using the conceptual daily rainfall-runoff model SIMHYD on 44 catchments in south-east Australia. The results indicate that the model averaging approach shows promise for estimating streamflow in ungauged catchments. The streamflow simulations are significantly better when model parameter sets are retained using the model averaging approach than when parameter values are estimated using the nearest neighbour and regression approaches. The results from the model averaging approach are also better than the cross-validation results in over 40 % of the catchments. It is likely that more detailed analyses of the choice of weights and descriptors of catchment similarity will lead to even better modelling results.
