AbstractsStatistics

Doutoramento em Engenharia Florestal e dos Recursos Naturais - Instituto Superior de Agronomia Forest ecosystems have been experiencing fast and abrupt changes in the environmental conditions, that can increase their vulnerability to extreme events such as drought, heat waves, storms, fire. Process-based models can draw inferences about future environmental dynamics, but the reliability and robustness of vegetation models are conditional on their structure and their parametrisation. The main objective of the PhD was to implement and apply modern computational techniques, mainly based on Bayesian statistics, in the context of forest modelling. A variety of case studies was presented, spanning from growth predictions models to soil respiration models and process-based models. The great potential of the Bayesian method for reducing uncertainty in parameters and outputs and model evaluation was shown. Furthermore, a new methodology based on a combination of a Bayesian framework and a global sensitivity analysis was developed, with the aim of identifying strengths and weaknesses of process-based models and to test modifications in model structure. Finally, part of the PhD research focused on reducing the computational load to take full advantage of Bayesian statistics. It was shown how parameter screening impacts model performances and a new methodology for parameter screening, based on canonical correlation analysis, was presented