Data assimilation with sensitivity-based particle filter: A simulation study with AquaCrop

Abstract

Crop growth simulation models are important components in agricultural management. Such models include parameters that should be calibrated locally, which can be achieved by assimilating observations using a particle filter. The number of model parameters is usually high while the number of observations is limited and adjusting parameters that are non-influential under the specific environmental conditions and growth stages should be avoided. This study suggests a novel particle filter-based framework in which sensitivity analysis (SA) is embedded in the filter so that at each data assimilation step only a subset of influential parameters is adjusted. The proposed framework was implemented in two synthetic study cases with the open-source AquaCrop model (v5.0a), assuming weekly observations of canopy cover and soil water content, and in some cases biomass. In the first case study, the adjustment of a subset of influential parameters identified by SA was compared with the adjustment of all candidate parameters, the impact of the SA screening threshold and the variance of the parameter perturbation included in the filter were investigated, and the addition of biomass measurements was examined. In the second case study, different irrigation treatments were implemented to demonstrate the impact of the growing conditions on the subset of parameters that could be calibrated. The performance of the proposed framework was evaluated by computing the root mean square error (RMSE) and normalized RMSE (NRMSE) of the states and parameters estimations, and of the final biomass and yield forecasts. Overall, all state predictions were very accurate. Estimation of the model parameters was more challenging and not all the parameters converged toward their true values. This result is not surprising considering the low sensitivity of certain parameters and correlations that exist within the model. Nonetheless, and more importantly, after the assimilation of relatively few observations, the model was able to forecast final biomass and yield quite accurately. Compared to standard data assimilation in which all parameters were adjusted, the SA-embedded filter performed better according to all the indicators considered:NRMSE of canopy cover and soil water content decreased from 2.4% to 1.3% and from 3.2% to 2.4%, respectively, NRMSE of parameter estimations decreased from 13.5% to 11.4%, and NRMSE of forecasted yield decreased from 5.9% to 5.2%. Overall, a relative improvement of 16% was obtained in the average NRMSE. Assimilation of additional biomass measurements improved only the ability of the model to forecast biomass but had no positive impact on yield forecasting. The next step should be to test such an approach thoroughly with experimental data.