Addressing initialisation uncertainty for end-to-end ecosystem models: application to the Chatham Rise Atlantis model

Abstract

Ecosystem models require the specification of initial conditions, and these initial conditions have some level of uncertainty. It is important to allow for uncertainty when presenting model results, because it reduces the risk of errant or non-representative results. It is crucial that model results are presented as an envelope of what is likely, rather than presenting only one instance. We perturbed the initial conditions of the Chatham Rise Atlantis model and analysed the effect of this uncertainty on the model’s dynamics by comparing the model outputs resulting from many initial condition perturbations. At the species group level, we found some species groups were more sensitive than others, with lower trophic level species groups generally more sensitive to perturbations of the initial conditions. We recommend testing for robust system dynamics by assessing the consistency of ecosystem indicators in response to fishing pressure under perturbed initial conditions. In any set of scenarios explored using complex end-to-end ecosystem models, we recommend that associated uncertainty analysis be included with perturbations of the initial conditions.

From data compilation to model validation: comparing three ecosystem models of the Tasman and Golden Bays, New Zealand

“…The lower trophic level species groups were generally more sensitive to perturbations of the initial conditions. While this corresponds well with higher variability expected at lower trophic levels in nature (Molinero et al, 2008), it was more likely caused by how we modelled the lower trophic levels (McGregor, Fulton & Dunn, 2020), as these were biomass pools with growth effectively instantaneous, and hence could be more responsive. Scallops and flatfish were exceptions as they were both modelled with age structure, and yet displayed high variability under historical fishing.…”

“…The alternative models using the Ecopath with Ecosim (EwE) and multi-species sizestructured model frameworks were developed as simplifications of TBGB_AM, and have been assessed with respect to TBGB_AM where it was possible to do so. TBGB_AM has been tested with respect to initialisation uncertainty, realised growth and mortality rates, variability from the assumed oceanographic variables, and connectivity analysis following the methods of McGregor et al (2019), andDunn (2020). All three models and associated R scripts are available on GitHub (McGregor, 2019).…”

“…An Atlantis model is run with a burn-in period followed by a model period. The purpose of the burn-in period is to allow initial fluctuations resulting from the initial state of the model to stabilise (McGregor, Fulton & Dunn, 2020). TBGB_AM consists of a 35-year burn-in period (1865-1899), followed by a model period from 1900 to 2014 (inclusive).…”

“…Initial conditions were perturbed for TBGB_AM following the methods of McGregor, Fulton & Dunn (2020). The resulting simulations were then used for assessing responses of TBGB_AM to historical fishing, thus allowing us to present these results as envelopes of plausibility rather than single trajectories.…”

Exploring the role of Northeast Atlantic cod in the Barents Sea food web using a multi‐model approach

“…The practice of considering model uncertainty in implemented ecosystem models therefore appears to be relatively rare. Several recent efforts have largely focused on uncertainty in parameter values, initial conditions and the process uncertainty that arises from natural variation (Bracis et al, 2020; Hansen et al, 2019; McGregor et al, 2020), whereas uncertainties about model structure have received less attention (Geary et al, 2020; Hill et al, 2007; Wildermuth et al, 2018).…”