Assessing lead-time for predicting wheat growth using a crop simulation model

Abstract

In order to use crop simulation models to predict crop yield, unobserved daily weather, an important input for crop models, must be forecast in some sense. Due to the chaotic nature of weather and the non-linear response of crop simulation models to weather input, this forecast weather cannot simply be a single weather series (e.g. average historical weather for the upcoming growing season), but must be an ensemble of weather series, incorporating site-specific climatic variability. To capture weather uncertainty, we used the LARS-WG stochastic weather generator to produce a probabilistic ensemble of weather series by mixing observed weather from the beginning of a season with stochastically generated (synthetic) weather for the remainder of the growing season. This ensemble was used with the crop simulation model Sirius to generate distributions of crop characteristics. Progressing through the growing season, as the proportion of synthetic weather in these ensembles decreased, the distribution means converged towards the true values, allowing us to make predictions with a high level of confidence before crop maturity. In this fashion, we analysed six sites with diverse climates in Europe and New Zealand, comparing lead-times for predicting different crop characteristics at various geographic locations. We demonstrated that that there is a large difference between lead-times amongst different crop characteristics at a single location, and that there is a large variation in lead-times for predicting selected crop characteristics between locations. Variation in climates places a quantifiable limit on our ability to make crop predictions using crop simulation models.