|Title||Parameter transferability across spatial and temporal resolutions in hydrological modelling|
|Author(s)||Melsen, L.A.; Teuling, A.J.; Torfs, P.J.J.F.; Zappa, M.; Mizukami, N.; Clark, M.; Uijlenhoet, R.|
|Event||EOS supplement, Fall Meeting of the American Geophysical Union, San Francisco, 2008-12-14/2008-12-18|
Hydrology and Quantitative Water Management
|Publication type||Abstract in scientific journal or proceedings|
|Abstract||Improvements in computational power and data availability provided new opportunities for hydrological modeling. The increased complexity of hydrological models, however, also leads to time consuming optimization procedures. Moreover, observations are still required to calibrate the model. Both to decrease calculation time of the optimization and to be able to apply the model in poorly gauged basins, many studies have focused on transferability of parameters. We adopted a probabilistic approach to systematically investigate parameter transferability across both temporal and spatial resolution. A Variable Infiltration Capacity model for the Thur basin (1703km2, Switzerland) was set-up and run at four different spatial resolutions (1x1, 5x5, 10x10km, lumped) and three different temporal resolutions (hourly, daily, monthly). Three objective functions were used to evaluate the model: Kling-Gupta Efficiency (KGE(Q)), Nash-Sutcliffe Efficiency (NSE(Q)) and NSE(logQ). We used a Hierarchical Latin Hypercube Sample (Vorechovsky, 2014) to efficiently sample the most sensitive parameters. The model was run 3150 times and the best 1% of the runs was selected as behavioral.
The overlap in selected behavioral sets for different spatial and temporal resolutions was used as indicators for parameter transferability. There was a large overlap in selected sets for the different spatial resolutions, implying that parameters were to a large extent transferable across spatial resolutions. The temporal resolution, however, had a larger impact on the parameters; it significantly affected the parameter distributions for at least four out of seven parameters. The parameter values for the monthly time step were found to be substantially different from those for daily and hourly time steps. This suggests that the output from models which are calibrated on a monthly time step, cannot be interpreted or analysed on an hourly or daily time step. It was also shown that the selected objective function affected the transferability. There was more overlap in parameter values between the time steps when the model was evaluated with the NSE(logQ) rather than with the KGE(Q) or the NSE(Q). This suggests that parameter transferability across temporal resolutions also depends on the time scale of the process under study.