Mikhail Semenov
Biomathematics & Bioinformatics, Rothamsted Research
Harpenden, Herts, AL5 2JQ, UK
Francisco J. Doblas-Reyes
ECMWF, Shinfield Park
RG2 9AX, Reading, UK
In order to predict crop yield in real time using crop simulation models, daily weather for
unobserved part of the growing season must be predicted in some sense. Seasonal forecasts up to
six months in advance, based on coupled ocean-atmosphere climate models, are now available at a
number of operational meteorological centres around the world. Seasonal forecasts are not suitable
directly for crop simulations, because of model biases and mismatch of spatial and temporal scales.
However, it is possible to utilise seasonal forecasts for crop modelling by constructing site-specific
daily weather using a stochastic weather generator linked with seasonal forecasts. For our study, we
use a subset of the DEMETER predictions, i.e. seasonal ensemble predictions from the ECMWF
GCM model. Daily weather time series, consisting of observed weather from the beginning of a
season and stochastically generated weather for the remainder of the season, were generated by the
LARS-WG stochastic weather generator. To assess usefulness of seasonal forecast, two sets of
scenarios were created based on seasonal forecast and historical climatology. Generated weather
was used as an input to the Sirius wheat simulation model to compute distributions of grain yield.
Two sites, one in Europe and one in New Zealand, were used in our analysis, where accuracy of
yield predictions was compared.