Predicting Runoff Signatures Using Regression and Hydrological Modeling Approaches

Accurate prediction of runoff signatures is important for numerous hydrological and water resources applications. However, there are lack of comprehensive evaluations of various approaches for predicting hydrological signatures. This study, for the first time, introduces regression tree ensemble approach and compares it with other three widely used approaches (multiple linear regression, multiple log-transformed linear regression, and hydrological modeling) for assessing prediction accuracy of 13 runoff characteristics or signatures, using a large data set from 605 catchments across Australia. The climate, in particular, mean annual precipitation and aridity index, has the most significant influence on the runoff signatures. Physical catchment attributes including forest ratio, slope, and soil water holding capacity also have significant influence (p < 0.05) on the runoff signatures. All four approaches can predict the long-term average and high flow signatures accurately. The regression approaches can also well predict majority of the other runoff signatures, with the Nash-Sutcliffe Efficiency larger than 0.60. The regression tree ensemble outperforms the two linear regressions in predicting signatures of flow dynamics. The hydrological models, calibrated to one specific objective criterion, cannot predict many of the runoff signatures, particularly those reflecting low flows and flow dynamics. This is because in most hydrological model applications, the simulations allow satisfactory predictions of long-term average and high flow signatures. In applications where a specific runoff signature is needed, regression relationships that directly relate that runoff signature to catchment attributes give the best predictions. Here the regression tree ensemble is overall best and offers significant potential, being able to predict most of the runoff signatures very well.