Uncertainty assessment, calibration and sensitivity analysis of process-based forest ecosystem computer simulation models

The use of process-based computer simulation models in forestry and the ecological sciences has advanced the need for tools to evaluate how well the models meet their intended purposes. Three aspects of model evaluation were addressed, uncertainty assessment, calibration and sensitivity analysis. Techniques relevant to these subjects were investigated and advanced. The methods were implemented using simple examples for demonstrative purposes and applied to a comprehensive evaluation of the forest ecosystem model PnET-II.; Bayesian synthesis (BYSN) was carried out as a tool for uncertainty assessment and calibration of model inputs. BYSN incorporates all available sources of information about the ecosystem of interest and uses the information to arrive at a joint distribution of model inputs and outputs. Marginal distributions were readily derived to identify credible regions of the parameter space where the actual system values would be expected to exist. The development of a calibration procedure using BYSN contributed an important tool for calibration of process-based models. A method for accounting for correlations between model inputs in BYSN was developed.; Sensitivity between model inputs and outputs was assessed using a global sensitivity assessment technique known as the Fourier amplitude sensitivity test (FAST). FAST was used to generate samples for a regression-based response surface model. Computation of FAST sensitivity coefficients (SIs) facilitated the assessment of global sensitivities and the development of error budgets for model outputs of interest. FAST generated correlation-free samples efficiently and the SIs compared favorably with an error budget obtained from the response surface model. A method of accounting for relationships between model inputs was developed. FAST results were used to identify inputs that had no effect on model outputs. As an outgrowth of the evaluation it was possible to identify some strengths and weakness of PnET-II and to suggest components of the model for which improvements can be made.