Regional crop yield forecasting using probabilistic crop growth modelling and remote sensing data assimilation

Information on the outlook of yield and production of crops over large regions is essential for government services dealing with import and export of food crops, for agencies playing a role in food relief, for international organisations with a mandate in monitoring the world food production and trade, as well as for commodity traders. In Europe, such information is provided by the MARS (Monitoring Agriculture with Remote Sensing) Crop Yield Forecasting System operated by the Joint Research Centre. An important component in the MARS Crop Yield Forecasting System is the so-called Crop Growth Monitoring System (CGMS). This system employs the WOFOST crop growth model to determine the influence of soil, weather and management on crop yield with a spatial resolution of 50×50 km grid. Aggregated CGMS results are used as predictors for crop yield at the level of EU member states. CGMS is being applied succesfully within the framework of the MARS crop yield forecasting system. Nevertheless, there are large uncertainties related to applying WOFOST over large areas such as poorly known sowing dates and soil parameters, application of irrigation and the effect of drought due to limited weather station density. This thesis focuses on developing and testing methods for quantifying and reducing uncertainty in crop model simulations with a focus on reducing the uncertainty related to drought. The uncertainty in crop model simulations is quantified through the variability within an ensemble of models, while it is reduced by combining crop model simulations with satellite-derived information through an ensemble Kalman filter (EnKF). A key aspect in this approach is that the uncertainty of the different components of the system can be estimated. The ultimate goal is to improve the accuracy and timeliness of regional crop yield forecasts. It was demonstrated that the uncertainty in the interpolated meteorological forcings is important, particularly the uncertainty in precipitation fields. Therefore, a method was developed to generate equiprobable realisations of precipitation inputs which can be used as input in the crop simulation model. It was demonstrated that the statistical properties of the precipitation field were reproduced reasonably well in the realisations, while the deviations from