Articles | Volume 14, issue 3
Hydrol. Earth Syst. Sci., 14, 447–458, 2010
https://doi.org/10.5194/hess-14-447-2010
Hydrol. Earth Syst. Sci., 14, 447–458, 2010
https://doi.org/10.5194/hess-14-447-2010

  08 Mar 2010

08 Mar 2010

Selection of an appropriately simple storm runoff model

A. I. J. M. van Dijk A. I. J. M. van Dijk
  • CSIRO Land and Water, Canberra, ACT, Australia

Abstract. An appropriately simple event runoff model for catchment hydrological studies was derived. The model was selected from several variants as having the optimum balance between simplicity and the ability to explain daily observations of streamflow from 260 Australian catchments (23–1902 km2). Event rainfall and runoff were estimated from the observations through a combination of baseflow separation and storm flow recession analysis, producing a storm flow recession coefficient (kQF). Various model structures with up to six free parameters were investigated, covering most of the equations applied in existing lumped catchment models. The performance of alternative structures and free parameters were expressed in Aikake's Final Prediction Error Criterion (FPEC) and corresponding Nash-Sutcliffe model efficiencies (NSME) for event runoff totals. For each model variant, the number of free parameters was reduced in steps based on calculated parameter sensitivity. The resulting optimal model structure had two or three free parameters; the first describing the non-linear relationship between event rainfall and runoff (Smax), the second relating runoff to antecedent groundwater storage (CSg), and a third that described initial rainfall losses (Li), but which could be set at 8 mm without affecting model performance too much. The best three parameter model produced a median NSME of 0.64 and outperformed, for example, the Soil Conservation Service Curve Number technique (median NSME 0.30–0.41). Parameter estimation in ungauged catchments is likely to be challenging: 64% of the variance in kQF among stations could be explained by catchment climate indicators and spatial correlation, but corresponding numbers were a modest 45% for CSg, 21% for Smax and none for Li, respectively. In gauged catchments, better estimates of event rainfall depth and intensity are likely prerequisites to further improve model performance.