|Ph.D Student||Jamal Alaa|
|Subject||A Framework for Improving the Predictions of Crop Growth|
Models using Real-Time Measurements
|Department||Department of Civil and Environmental Engineering||Supervisor||Professor Raphael Linker|
|Full Thesis text|
Although simulation models are often used for state estimations, these models are rarely accurate due to uncertainties in boundary conditions, initial conditions and/or model parameters. Field measurements, although partial and inaccurate, can be used to improve the state estimations via so-called data assimilation. Some of the most important methods for data assimilation include the Kalman filter (KF) and ensemble Kalman filter (EnKF) for non-linear systems. Filter divergence is a well-known phenomenon which ultimately causes the estimation of the state vector to drift away from the true state vector. Covariance inflation can be applied to alleviate this phenomenon. Existing covariance inflation methods are often based on pre-determined stationary inflation amplitude or involve highly influential tuning parameters. In the first part of the present study, a novel approach that uses statistically-based confidence intervals of the measurements to estimate time-varying inflation factors was developed. The main advantage of this method is that it does not involve user-defined tuning parameters. The proposed method was tested on a soil hydrology case study and the results were compared to those obtained with methods from the literature. Our proposed method showed better estimations and faster convergence toward the true state values. In particular, state estimation evolved smoothly whereas the literature method exhibited a somewhat erratic behavior, most likely due to underlying assumptions of the method not being met.
The assumptions on which EnKF is based are rarely strictly met and in such cases particle filter (PF) is a more suitable alternative. Various modifications of the PF have been proposed in the literature, among them a recent method that combines Markov Chain Monte Carlo and evolutionary operators with PF. In the second part of the present study, a new data assimilation procedure inspired from this approach was developed. The main novelty of the proposed approach is that the state and parameters are estimated simultaneously rather than sequentially as in the aforementioned work. The method was tested on two simulation examples involving a crop model, and was compared to other existing methods. The proposed method showed robust performance in predicting states and parameters, and led to improved predictions of canopy cover, water content, final biomass and yield. However, not all crop and soil parameters were estimated accurately, which was due to the fact that some of these parameters influenced the predictions only marginally. This observation led to combining data assimilation with sensitivity analysis in the third part of the thesis.
Performing sensitivity analysis prior to data assimilation reduced the number of parameters involved in the updating step, leading to faster convergence and overall improved predictions. The combined method was tested with a large synthetic experiment and data collected in a small-scale experiment. In the synthetic case study, the method showed superior estimations of the current states and future predictions of the states. In the experimental case study, adding global sensitivity analysis to the PF scheme reduced significantly the number of parameters adjusted without affecting the overall performance of the data assimilation procedure.