Ecological models can take on a variety of forms (e.g., data-driven, process-driven) and can be used for a variety of purposes (e.g., prediction, forecasting, increasing understanding, decision-making). However, despite these differences, all ecological models consist of a functional relationship between a set of model inputs and parameters and one or more model outputs. All of these components of ecological models (i.e., inputs, parameters, functional form) are subject to considerable uncertainty (e.g., data, natural variation, a lack of knowledge and understanding). In order to be able to use ecological models with confidence, these uncertainties need to be addressed explicitly. This can be achieved by using SA and UA. Both attempt to propagate uncertainties in model inputs and parameters to the model output(s). However, in SA, uncertainties in model inputs and parameters are represented by likely ranges. In contrast, in UA, the uncertainty in each of the inputs and parameters is first characterized to some degree (often by a PDF), enabling probabilistic information, such as confidence limits on predictions and forecasts, to be obtained.
SA can be used to assess model validity, as well as the relative sensitivities of model outputs to model inputs/ parameters. Due to the highly nonlinear nature of ecological models, and the strong possibility of interactions between model inputs and parameters, sampling-based SA methods (e.g., Monte Carlo methods) are most appropriate in an ecological modeling context. Stratified (e.g., Latin hypercube) or transformation-based (e.g., FAST) sampling regimes can be used if computational efficiency is an issue.
UA can be used to determine the moments and/or distributions of model outputs and can therefore be used to obtain confidence limits on predictions, as well as a range of risk-based performance measures. If no calibration data are available, uncertainties in model inputs can be determined based on experience, values from the literature, or available input data, and propagated through the model using Monte Carlo methods or first- or second-order approximations. When calibration data are available, more sophisticated methods, such as GLUE or MCMC, have to be used to ensure that the generated model outputs are in agreement with the available data, and output distributions have to be generated by sampling from the parameter/input vectors obtained as part of the GLUE/MCMC analysis.
Was this article helpful?