Ensemble modelling of the hydrological impacts of land use change

Abstract

Ensemble modelling, whereby predictions from several models are pooled in an attempt to improve prediction accuracy, has often been used in the climate and atmospheric sciences, but until recently, has received little attention in hydrology. One of the key aims of the ensemble approach is to reduce uncertainty in the modelled predictions. This paper reports on a project to compare predictions from a range of catchment models applied to a mesoscale river basin in central Germany and to produce ensemble predictions of the effects of several projected land use changes. The models encompass a large range in inherent complexity and input requirements. In approximate order of decreasing complexity, they are DHSVM, MIKE-SHE, TOPLATS, WASIM-ETH, SWAT, PRMS, SLURP, HBV, LASCAM and IHACRES. Overall, the simpler models tend to perform better in both calibration and validation, but while all models tend to show improved performance during the lessextreme validation period, this improvement is greatest for some of the more complex models. Despite the disparity in model performance, three ensemble predictions made up of various combinations of the 10 model predictions outperform all of the individual models. In calibration, the ensemble based on a multi-variable regression of all models provides the best predictions, but its prediction accuracy declines to a greater extent than all of the models in terms of both its bias and Nash-Sutcliffe efficency when used in the validation period. In the validation period, the best predictions are provided by an ensemble consisting of the daily median model predictions. The predictions of this median ensemble also improve more between calibration and validation than any of the other models, thus indicating its robustness. The calibrated models are applied to three land use change scenarios. In the scenarios, the projected patterns of land use are based on assumed average field sizes of 0.5 ha, 1.5 ha and 5.0 ha, respectively. These contrast with a current average field size of about 0.7 ha. There is broad agreement among the models on the expected hydrological change (Figure 1). This, coupled with the validation success of the mean and median ensembles, suggests that we can predict with some confidence the direction and magnitude of streamflow changes associated with the three scenarios. (Graph Presented).