| The reliability of aquatic biogeochemical models is inversely proportional to the amount of uncertainty associated with model predictions. This uncertainty is a function of error arising from model specification, parameter estimation, and computational methods. Here, the role of observations (data) and a priori knowledge in mitigating the first two sources of error is explored, with an emphasis on the identification of the optimal level of model complexity for a given situation. Information-theoretic model selection techniques are applied to a set of simple one-dimensional biogeochemical models describing chlorophyll in Trout Lake using a large field data set; results demonstrate that even relatively modest levels of complexity are often statistically unsupportable by available data. This is one of the first applications of information theory to development of aquatic biogeochemical models. Since the scope of problems being addressed by environmental models often necessitates a greater level of complexity than can be mathematically justified, the application of formal parameter estimation techniques (rather than manual tuning based solely on a priori knowledge) is essential, yet is often prohibited by computational costs. A novel method is developed in which the vertical hydrodynamics of a three-dimensional model is emulated in one-dimensional space, allowing biogeochemical parameter values to be optimized using the available data at a reasonable computational expense. It is also demonstrated that a simple model formulation can, in some cases, provide a more useful tool than a more detailed representation of biogeochemical dynamics, due to the greater uncertainty in both model structure and parameterization associated with the more complex model. A simple model of biogeochemical cycling is developed for Lake Superior that yields a better fit to existing data than do more mechanistically detailed formulations. Using this model, gross primary production (GPP) in Lake Superior is estimated to be ∼10 Tg C yr-1, an amount equal to 12--77% of community respiration (CR) in the lake. The model also suggests that the deep chlorophyll maximum in Lake Superior may be explained primarily by algal shade adaptation. Finally, even in the case where both constraining data and adequate knowledge of system structure are insufficient, simple models can still be formulated to test hypotheses, though the amount of uncertainty associated with such models is high. A series of simple model structures are calibrated using historical Lake Superior nitrate data in order to identify the possible historical sequence of events that caused a precipitous rise in nitrate concentrations during the previous century. The model suggests that atmospheric deposition alone was likely insufficient to result in the observed increase, and that either loading was elevated or burial was depressed mid-century. |
