Ecosystem states: Creating a data-derived, ecosystem-scale ecological response model that is explicit in space and time

Increasing difficulties associated with balancing consumptive demands for water and achieving ecological benefits in aquatic ecosystems provide opportunities for new ecosystem-scale ecological response models to assist managers. Using an Australian estuary as a case study, we developed a novel approach to create a data-derived state-and-transition model. The model identifies suites of co-occurring birds, fish, benthic invertebrates and aquatic macrophytes (as ‘states’) and the changing physico-chemical conditions that are associated with each (‘transitions’). The approach first used cluster analysis to identify sets of co-occurring biota. Differences in the physico-chemical data associated with each state were identified using classification trees, with the biotic distinctness of the resultant statistical model tested using analysis of similarities. The predictive capacity of the model was tested using new cases. Two models were created using different time-steps (annual and quarterly) and then combined to capture both longer-term trends and more-recent declines in ecological condition. We identified eight ecosystem states that were differentiated by a mix of water-quantity and water-quality variables. Each ecosystem state represented a distinct biotic assemblage under well-defined physico-chemical conditions. Two ‘basins of attraction’ were identified, with four tidally-influenced states, and another four independent of tidal influence. Within each basin, states described a continuum of relative health, manifest through declining taxonomic diversity and abundances. The main threshold determining relative health was whether freshwater flows had occurred in the region during the previous 339days. Canonical analyses of principal coordinates tested the predictive capacity of the model and demonstrated that the variance in the environmental data set was well captured (87%) with 52% of the variance in the biological data set also captured. The latter increased to >80% when long- and short-term biological data were analysed separately, indicating that the model described the available data for the Coorong well. This approach thus created a data-derived, multivariate model, where neither states nor transitions were determined a priori. The approach did not over-fit the data, was robust to patchy or missing data, the choice of initial clustering technique and random errors in the biological data set, and was well-received by local natural resource managers. However, the model did not capture causal relationships and requires additional testing, particularly during future episodes of ecological recovery. The approach shows significant promise for simplifying management definitions of ecological condition and, via scenario analyses, can be used to assist in manager decision-making of large, complex aquatic ecosystems in the future.