Modelling the effects of management on population dynamics: some lessons from annual weeds

摘要

1. Agricultural and invasive weeds are major threats to managed and natural ecosystems, costing billions of dollars annually. Models for arable and invasive weed population dynamics can contribute to the management of problem species through generating predictions of population densities, and how those are likely to respond to changing management. Frequently, however, little attention is paid to quantifying the errors in estimates of model parameters and how these may affect the robustness of model predictions. 2. Most weed models are parameterized using a combination of field-estimated and literature-derived parameters. Only rarely are estimates of error available for all parameters, despite the fact that most parameters will be subject to considerable error. 3. Close to extinction boundaries, predictions of population densities may be highly sensitive to small errors in model parameters. However, because of their generally high reproductive capacities, many weeds may occur at high and economically significant densities while close to extinction. Consequently, models may be numerically unstable at ecologically realistic densities. 4. We review methods of dealing with parameter uncertainty in weed modelling. We stress that it is important to recognize that many models may be structurally incorrect and all stabilizing mechanisms may not have been identified. Also, alternative model forms have seldom been explored, although a variety of alternatives to conventional difference equations exist. 5. Synthesis and applications. Only when different management interventions have greatly contrasting effects on population sizes are most models of weed populations likely to provide qualitatively correct predictions of their effects. Quantitative predictions from demographic models will usually be subject to large errors. Modelling methods that account for spatial heterogeneity and other stabilizing effects may yield more accurate predictions; however, their parameterization will often require approaches for data collection very different from those currently used.