Reduction of predictive uncertainty in estimating irrigation water requirement through multi-model ensembles and ensemble averaging
- 1Institute for Landscape Ecology and Resources Management (ILR), Research Centre for BioSystems, Land Use and Nutrition (IFZ), Justus Liebig University Giessen, Heinrich-Buff-Ring 26, 35390 Giessen, Germany
- 2School of GeoSciences and National Centre for Earth Observation, University of Edinburgh, Edinburgh, UK
- 3Climate Change Research Centre and ARC Centre of Excellence for Climate System Science, University of New South Wales, Sydney, New South Wales, Australia
- 4CSIRO Land and Water, GPO Box 1666, Canberra, ACT 2601, Australia
Abstract. Irrigation agriculture plays an increasingly important role in food supply. Many evapotranspiration models are used today to estimate the water demand for irrigation. They consider different stages of crop growth by empirical crop coefficients to adapt evapotranspiration throughout the vegetation period. We investigate the importance of the model structural versus model parametric uncertainty for irrigation simulations by considering six evapotranspiration models and five crop coefficient sets to estimate irrigation water requirements for growing wheat in the Murray–Darling Basin, Australia. The study is carried out using the spatial decision support system SPARE:WATER. We find that structural model uncertainty among reference ET is far more important than model parametric uncertainty introduced by crop coefficients. These crop coefficients are used to estimate irrigation water requirement following the single crop coefficient approach. Using the reliability ensemble averaging (REA) technique, we are able to reduce the overall predictive model uncertainty by more than 10%. The exceedance probability curve of irrigation water requirements shows that a certain threshold, e.g. an irrigation water limit due to water right of 400 mm, would be less frequently exceeded in case of the REA ensemble average (45%) in comparison to the equally weighted ensemble average (66%). We conclude that multi-model ensemble predictions and sophisticated model averaging techniques are helpful in predicting irrigation demand and provide relevant information for decision making.