Data assimilation in the Digital Future Farm: from prototypes to operational use

Within the DFF modelling framework, there are several models for estimating production of crops and grasslands, estimating use of water and nutrients, calculating feed requirements of livestock and their production of beef and milk as well as manure. However, models are only a simplification of reality and often their predictions deviate considerably from the true situation. To correct models for biases we can use data assimilation (DA) techniques which use observations to make adjustments to the model simulations in order to get the simulation “back on track”. DA techniques are commonly used, for example in todays weather forecasting models where they have been very successful in improving weather forecasts.

In 2020, we have implemented a commonly used DA technique within the DFF modelling framework: the Ensemble Kalman Filter (EnKF). The principle behind the EnKF is that it assumes that both models and observations are uncertain. Therefore it uses a weighted estimate of both the model prediction and the observation to estimate the true state of the simulated system. Uncertainty in the model predictions is represented by an ensemble of model results.

Our first results using the WOFOST model (for potato) and the GRAS2007 model (for grasslands) using observed vegetation canopy variables (LAI, height) have demonstrated that the use of the EnKF reduces the error on the model predictions. This also works for states that have not directly observed (total biomass, yield). Therefore, in 2021 we aim to repeat this exercise for more crops and situations and operationalize the EnKF in the DFF framework by ingesting satellite observations that are routinely available.