Performance of ten rainfall–runoff models applied to an arid catchment in Oman

When selecting a rainfall–runoff model for application to an arid region, the literature dictates the need to consider the spatial features of rainfall and the variability and non-linearity of losses, and to match model complexity to the availability and quality of data. In light of this, the metric-conceptual IHACRES model is applied to hourly data from a 734 km2 catchment in Oman, using a semi-distributed representation of rainfall input. Sensitivity analysis is used to guide reduction of the model from a 9-parameter version to simpler versions. The performances of the alternative versions, for predicting flow volumes and peaks at the catchment outlet, are inter-compared. Performances are also compared with those achieved using lumped versions of IHACRES, a physics-based model and a 2-parameter regression model. For peak flows, a 2-parameter non-linear loss model with 2-parameter linear routing, applied in semi-distributed mode, achieves the best overall performance. For flow volumes, the same model was preferred although the routing component was not required. The principal reasons for the success of these models are thought to be their parsimony, representation of spatial rainfall, and ability to compensate for systematic rainfall and flow observation errors. Extra performance is achieved by using a score-based calibration criterion, which is more robust to extreme errors than the fit-based criteria. Although the best performances are poor, with an average absolute relative error across events of 53% for flow peaks and 36% for flow volumes, this is not disappointing compared to other applications of this type. Prediction uncertainty is high due to variability of effective parameter values over events, and uncertainty analysis must explicitly represent this variability in order to explain the observations.