Allard de Wit, Rob Knapen, Eliya Buyukkaya and Sander Janssen
Wageningen Environmental Research
January 2021 The ensemble Kalman filter (EnKF) is a proven methodology for assimilating external observations in crop models. For example, observations of LAI (leaf area index) and soil moisture from Copernicus (Satellite) missions can be used to adjust crop models taking field-specific conditions into account. The EnKF requires an ensemble of models to estimate model uncertainty and updates the model state whenever an observation is available weighting the uncertainty in the model ensemble and the observation.
A data assimilation experiment using the EnKF applying field experimental data demonstrates that observed LAI can be used to adjust the model, but without corrections the model drifts away from the observations as the growing season progresses (Figure 1). This problem, which is called “filter divergence”, is well known for crop models that tend to underestimate uncertainty in their predictions as the growing season progresses.
Figure 1. Assimilating potato leaf area index in the WOFOST cropping system model using the EnKF.
A solution to the problem of filter divergence can be to artificially add some uncertainty to the model predictions using a “variance inflation factor” (VIF). Using such an approach requires some careful tuning of the VIF value but ensure that the crop model keeps responsive to external observations. The latter is particularly important for including external effects that are not including in the model itself, such as the impact of lack of nutrients or pest and disease. Applying a correction factor to inflate the variance helps in adjusting the model to the observations (figure 2) and also helps decrease the error on the predicted yield (figure 3).
As part of the CYBELE project, we will test the methodology on more fields and crops in the coming months in order to further optimize the algorithm and improve performance.
Figure 2. Assimilating potato leaf area index in the WOFOST cropping system model using the EnKF including a variance inflation factor.
Figure 3. Error on predicted potato yield for the standard run, the ensemble only and two variants of data assimilation with the EnKF.