- This event has passed.
PhD defense – Fenni Kang
28 May 2013 @ 14 h 00 - 17 h 00
Title: Plant growth models and methodologies adapted to their parameterization for the analysis of phenotypes
Abstract: Plant growth models aim at describing the interaction between the growth of plants and their environment. Ideally, model parameters are designed to be stable for a wide range of environmental conditions, and thus to allow characterizing genotypes. They offer new tools to analyze the genotype X environment interaction and they open new perspectives in the process of genetic improvement. Nevertheless, the construction of these models and their parameterization remain a challenge, in particular because of the cost of experimental data collection. In this context, the first contribution of this thesis concerns the study of plant growth models. For sunower (Helianthus annuus L.), the model SUNFLO [Lecoeur et al., 2011] is considered. It simulates the plant phenology, morphogenesis and photosynthesis under abiotic stresses. An extension of this model is proposed: this new SUNLAB model adapts into SUNFLO a module of biomass allocation to organs, using the source-sink concepts inspired by the GREENLAB model [De Reffye and Hu, 2003]. For maize (Zea mays L.), the CORNFLO model, based on the same principles as SUNFLO, was also studied. These models helps discriminating genotypes and analyzing their performances. On the other hand, in order to parameterize these models, an original methodology is designed, adapted to the context of plant variety improvement by breeders. The MSPE methodology (“multi-scenario parameter estimation”) uses a limited number of experimental traits but in a large number of environmental configurations for the parameter estimation by model inversion. Issues including identifiability, sensitivity analysis, and the choice of optimization methods are discussed. The influences of environmental scenarios amount on the model predictive ability and on estimation error are also studied. Finally, it is demonstrated that selecting scenarios in different environmental classes (obtained by data clustering methods) allows to optimize the multi-scenario parameter estimation performances, by reducing the required number of scenarios.