Reconstruction of ecosystem flows using inverse methods:: how well do they work?

Inverse modelling is used to synthesize multivariate observations from marine and freshwater ecosystems into descriptions of mass flows among ecological and biogeochemical components. However, the conditions that affect the accuracy of these analyses remain poorly understood. In particular, it is suspected that the steady-state assumption often used in these analyses and the flow minimization principle that underlie inverse modelling introduce distortions into the reconstructions of ecosystem flows, but these potential biases have not been quantitatively investigated. Simulated inverse experiments were conducted to shed some light on these issues. In these experiments, inverse analyses are run on 'artificial' observations generated from a mechanistic ecological-biogeochemical model. The simulated experiments indicate that the steady-state assumption has little impact on the accuracy of inverse reconstructions of ecosystem flows. Inverse analyses run on observations from simulations of transient states are as accurate as analyses ran on observations from simulations at steady state. The accuracy of inverse reconstructions is related to structural and dynamic features of the ecosystem. Inverse reconstructions on simulated ecosystems with weak nutrient recycling (dependent mostly on external nutrients) or with simple food webs show little bias. Reconstructions of simulated ecosystems with strong recycling or complex food webs show significantly more bias, with a tendency to overestimate small flows and to underestimate large flows. Despite these biases, inverse reconstructions were successful at detecting changes in flow structure associated with changes in simulated ecosystem properties. The simulations also indicate that the inverse analyses based on simultaneous accounting of more than one currency (e.g. carbon+nitrogen) should be preferred over analyses based on balancing only one currency (e.g. carbon or nitrogen). (C) 2003 Elsevier Science B.V. All rights reserved.