Bayesian methods for predicting LAI and soil moisture

The prediction errors of crop models are often important due to uncertainties in the estimates of initial values of the states, in parameters, and in equations. The measurements needed to run the model are sometimes not numerous or known with some uncertainty.
In this paper, two Bayesian filtering methods were used to update the state variable values predicted by MiniSTICS model. The chosen state variates were the LAI (Leaf Area Index) of a wheat crop (Triticum aestivum L.) and the corresponding moisture content of two soil layers (0-20 cm and 30-50 cm). These state variates were estimated simultaneously with several parameters. The assessed filtering methods were the centralized Particle Filtering (PF) and the Variational Bayesian Filtering (VF). The former is known to be sensitive to the number of particles while the latter yields an optimal choice of the sampling distribution over the state variable by minimizing the Kullback-Leibler divergence. In fact, variational calculus leads to a simple Gaussian sampling distribution whose parameters (estimated iteratively) depends on the observed data. On basis of a case study, the VF method was found more efficient than the PF method. Indeed, with the VF, the Root Mean Square Error (RMSE) of the three estimated states was smaller and the convergence of the all parameters was ensured.