Comparison of multi-model and multi-donor ensembles for regionalisation of runoff generation using five lumped rainfall-runoff models

Five lumped, conceptual rainfall-runoff models are calibrated for 240 gauged catchments in southeastern Australia. Climate input to the models is distributed at ~25 km grid cells and the catchments range in size from 50 to 2000 km. Each of the models is calibrated on each of the 240 catchments. Each catchment is then simulated using parameters sets calibrated for the nearest neighbouring catchment. Model predictions are assessed using the daily Nash-Sutcliffe efficiency and the volume bias. The results demonstrate that whilst an increasing number of optimisable parameters leads to increased calibration performance (when assessed using metrics based on the sum of squared residuals), the reverse is true for a large proportion of catchments in cross-verification using parameters from one donor catchment. This reversal, however, does not persist for the multi-donor averages, where the more highly parameterised models typically have the best performance. A weighted average of the five models (weighted by calibration performance) is shown to yield better calibration predictions than an unweighted average, but in cross-verification there is little difference between the two. This suggests that the relative calibration performances of different models in a donor catchment are not necessarily good indicators of how well the models will contribute to prediction in a neighbouring catchment. Five-member multi-donor ensembles of each individual model, weighted by distance, are superior to using a raw average, and both unweighted and weighted multi-donor ensembles are superior to the respective singledonor models. This indicates that while there is useful information delivered to an ensemble by the fifth nearest catchment, the value of this information is not as significant as the information from the nearest catchment. Further investigation using a multi-donor ensemble approach indicates that the optimum number of catchments to include in a spatial ensemble is five or six, and that such an ensemble can lead to considerable improvement in runoff predictions in ungauged catchments. We show that for the set of models and catchments used, the multi-donor approach using a single rainfall-runoff model is superior to the multi-model approach, but that a combination of the two approaches yields the best overall predictions. A five-member unweighted multi-model ensemble is shown to give regionalised predictions that are commensurate with the typical five-member unweighted multi-donor ensemble, but when the multi-model ensemble is weighted by donor calibration performance, its predictions are poorer than each of the multidonor models weighted by distance from the target catchment. Nonetheless, the best predictions assessed in this study are those of a multi-model multi-donor ensemble that combines the weighted averaging methods of both combinations.