The effects of model structure and complexity on the behaviour and performance of marine ecosystem models

Despite increasing use of ecosystem models, the effects of model structure and formulation detail on the performance of these models is largely unknown. Two biogeochemical marine ecosystem models were constructed as the foundation of a study considering many aspects of model simplification. The models use a trophic web that is resolved to the level of functional groups (feeding guilds), and includes the main pelagic and benthic guilds from primary producers to high-level predators. Both models are process based, but the Integrated Generic Bay Ecosystem Model (IGBEM) is highly physiologically detailed, while Bay Model 2 (BM2) uses simpler general assimilation equations. Both models compare well with real systems under a wide range of eutrophication and fishing schemes. They also conform to general ecological checkpoints and produce spatial zonation and temporal cycles characteristic of natural systems. The performance of IGBEM is not consistently better than that of BM2, indicating that high levels of physiological detail are not always required when modelling system dynamics. This was reinforced by a section of the study that fitted BM2, IGBEM and an existing ecosystem model (ECOSIM) to Port Phillip Bay. The predictions of all three models lead to the same general conclusions across a range of fishing management strategies and scenarios for environmental change. Models that are less resolved or use simpler formulations have lower computational demands and can be easier to parameterise and interpret. However, simplification may produce models incapable of reproducing important system dynamics. To address these issues simplified versions of BM2 and IGBEM were compared to the full models to consider the effects of trophic complexity, spatial resolution, sampling frequency and the form of the grazing and mortality terms used in the models on the performance of BM2 and IGBEM. It was clear in each case that some degree of simplification is acceptable, but that using models with very little resolution or very simplistic linear grazing and mortality terms is misleading, especially when ecosystem conditions change substantially. The research indicates that for many facets of model structure there is a non-linear (humped) relationship between model detail and performance, and that there are some guiding principles to consider during model development. Developmental recommendations include using a sampling frequency of 2 —4 weeks; including enough spatial resolution to capture the major physical characteristics of the ecosystem being modelled; using quadratic mortality terms to close the top trophic levels explicitly represented in the modelled web; aggregating species to the level of functional groups when constructing the model's trophic web, but if further simplification of the web is necessary then omission of the least important groups is preferable to further aggregation of groups; giving careful consideration to the grazing terms used, as the more complex lolling type responses may be sufficient; and if an important process or linkage is not explicitly represented in the model, or is poorly known, then a robust empirical representation of it should be included. The work presented here also has implications for wider ecological topics (e.g. the stability-diversity debate) and management issues. Consideration of the effects of trophic complexity on model performance under a range of environmental conditions supports the ecological "insurance hypothesis", but not the existence of a simple relationship between diversity and stability. The biological interactions captured in the web are a crucial determinant of ecosystem and model behaviour, but simple aggregate measures such as diversity, interaction strength and connectance are not. Similarly, the work on the effects of spatial resolution on model performance indicates that spatial heterogeneity is a crucial system characteristic that contributes to many of the emergent properties of the system. The comparison of the full models with each other, and with ECOSIM, leads to five general conclusions. First, shallow enclosed marine ecosystems react more strongly to eutrophication than to fishing. Second, a selected set of indicator groups can signal and characterise the major ecosystem impacts of alternative management scenarios and large-scale changes in environmental conditions. Third, policies focusing on the protection of a small sub-set of groups (especially if they are concentrated at the higher trophic levels) can fail to achieve sensible ecosystem objectives and may push systems into states that are far from pristine. Fourth, multispecies and ecosystem models can identify potential impacts of anthropogenic activities and environmental change that a series of single species models cannot. Finally, and most importantly, the use of a single "ultimate" ecosystem model is ill advised, but the comparative and confirmatory use of multiple "minimum-realistic" models is very beneficial.