Abstract: In this paper, an Improved Particle Filtering (IPF) based on minimizing Kullback-Leibler divergence will be proposed for biomass prediction of a wheat crop model in the context of climate change including heat and drought stresses.In a first stage, the performances of the proposed technique will be compared with those of the extended Kalman filter (EKF), unscented Kalman filter (UKF), Particle filter (PF). In a second stage, the state estimation techniques EKF, UKF, PF and IPF will be used for updating prediction of the model in order to predict winter wheat biomass, in specific field conditions, during several contrasted weather conditions. In a third stage, the effects of practical challenges on the performances of the state estimation algorithms will be assessed. Such practical challenges include the effect of measurement noise on the estimation performances and the measurement frequency of state variables.The first results show that the UKF provides a higher accuracy than the EKF due to the limited ability of EKF to accurately estimate the mean and covariance matrix of the estimated states through lineralization of the nonlinear process model. The results also show that the IPF provides a significant improvement over PF because, unlike the PF which depends on the choice of sampling distribution used to estimate the posterior distribution, the IPF yields an optimum choice of the sampling distribution, which also accounts for the observed data. For all techniques, the practical challenges affect the estimation accuracy as well as the convergence of the estimated states and parameters. However, the IPF can still provide both convergence as well as accuracy over other estimation methods. These advantages are precious in presence of high climate stresses.

Abstract: Crop models used as parts of integrated assessments often need to be run on regional, national and global scales. Calibration is an important step in the application procedure, yet on scales like this the process needs to be simplified in order to meet data requirements and computational limits. The question arises if a model calibrated in such a “simple” fashion still performs adequately at field scale, and if parameters not calibrated in the process can nevertheless be used with some confidence in later stages of the assessment. To answer this question, we applied the crop model EPIC to the simulation of sugarcane in Sao Paulo, Brazil. We once calibrated the model using Bayesian calibration to data on yield, aboveground biomass, and root weight measured in four years on two field trials in Sao Paulo. For the second calibration we used a simplified approach and calibrated the model only to FAOSTAT yield data for the whole of Brazil. Both calibrated models were applied to the simulation of stalk yield, aboveground biomass and root weight on a third field trial, and to the simulation of mean yields in Sao Paulo. The results showed that both models were able to adequately depict yields on both scales, but that the model calibrated to only national yield data was not able to accurately simulate root biomass, and to a lesser degree aboveground biomass. We conclude that a simplified calibration performs adequately on both scales, but that non-calibrated parameters may only be used with caution.