A comparison of Bayesian and classical methods for parameter estimation in greenhouse crop models

I.L. López-Cruz, A. Ruiz-García, E. Fitz-Rodríguez, R. Salazar-Moreno, A. Rojano-Aguilar
Parameter estimation in crop growth models is useful for improving predictions. However, some issues related to the improvement on the predictions accuracy by model calibration still need to be addressed. Additionally, there are several unanswered practical questions. For instance, the issue of what data should be used in the calibration of the model is still an open question. There is no general consensus on what goodness-of-fit criterion should be used. Also, researchers do not know if either sequential or simultaneous calibration should be preferable. Furthermore, it is not clear which and how many parameters should be adjusted and how the information coming from the individual processes and overall system behaviour should be combined for improving the parameter estimation. Two paradigms are identified regarding the calibration of crop growth models: classical and Bayesian. Even though that the Bayesian approach has several advantages over the classical method, only few studies have been carried out on crop growth models and there are no studies at all in case of greenhouse crop models. A comparison between the classical procedure of nonlinear least squares and two Bayesian parameter estimation methods was carried out. A two state variable greenhouse crop model was used in order to show the strengths and weakness of both paradigms. Firstly, the lettuce model was calibrated by a well-established nonlinear least squares procedure. Secondly, for the Bayesian method, a prior parameter distribution was specified for some model parameters. After that, the posterior parameter distribution was computed by using the Bayes' theorem. The second step implies the use of a numerical solution due to the complexity of greenhouse crop model. Thus, both the generalized likelihood uncertainty estimation (GLUE), which can be considered as an importance sampling method and the Metropolis-Hastings algorithms were implemented. Results were similar between both paradigms.
López-Cruz, I.L., Ruiz-García, A., Fitz-Rodríguez, E., Salazar-Moreno, R. and Rojano-Aguilar, A. (2017). A comparison of Bayesian and classical methods for parameter estimation in greenhouse crop models. Acta Hortic. 1182, 241-248
DOI: 10.17660/ActaHortic.2017.1182.29
https://doi.org/10.17660/ActaHortic.2017.1182.29
greenhouse lettuce crop, nonlinear least squares, generalized likelihood uncertainty estimaton (GLUE), importance sampling, Metropolis-Hastings algorithm
English

Acta Horticulturae