Prediction of chlorophyll A concentrations in Lake Okeechobee

Lake Okeechobee water quality data from 1980 to 1993 were used to develop a model for predicting chlorophyll a concentrations. Data collected by the South Florida Water Management District (SFWMD) and the University of Florida included stage, temperature (TEMP), total nitrogen (TN), total phosphorous (TP), wind velocity and chlorophyll a. Three modeling approaches to the objective of prediction of chlorophyll a concentrations are developed in this work: (1) a general linear model, (2) a generalized additive model and (3) nonparametric regression with Bayesian variable selection which incorporates non-additive interactions among predictors.; In the initial analysis, an ordinary least squares model was fitted to predict log{dollar}sb{lcub}10{rcub}{dollar} chlorophyll a concentrations ({dollar}mu{dollar}g L{dollar}sp{lcub}-1{rcub}{dollar}) as a function of log{dollar}sb{lcub}10{rcub}{dollar} TP (mg L{dollar}sp{lcub}-1{rcub}{dollar}), log{dollar}sb{lcub}10{rcub}{dollar} TN (mg L{dollar}sp{lcub}-1{rcub}{dollar}), log{dollar}sb{lcub}10{rcub}{dollar} stage (ft msl), log{dollar}sb{lcub}10{rcub}{dollar} temperature (deg C) and date. This model provides prediction at a finer spatial scale than the mechanistic model developed by SFWMD by fitting different slopes and intercepts for each of five previously described ecological zones of Lake Okeechobee. The early mechanistic models used to predict chlorophyll a predicted only whole lake average concentrations. Residuals analysis of this model suggested nonlinear relationships between the chlorophyll a response and some of the predictors, which led to the use of generalized additive models.; In the second phase of the analysis, a generalized additive model (GAM) was fitted to predict log{dollar}sb{lcub}10{rcub}{dollar} chlorophyll a concentrations using the same predictors and ecological zone scheme used for the ordinary least squares model. GAMs suggest that nonlinear relationships exist, notably the log{dollar}sb{lcub}10{rcub}{dollar} TP-log{dollar}sb{lcub}10{rcub}{dollar} chlorophyll a and log{dollar}sb{lcub}10{rcub}{dollar} Stage-log{dollar}sb{lcub}10{rcub}{dollar} chlorophyll a relationships. For TP, the additive functions in some of the ecological zones indicate a decreasing log{dollar}sb{lcub}10{rcub}{dollar} chlorophyll a concentration when log{dollar}sb{lcub}10{rcub}{dollar} TP increases above a certain (high) level, suggesting light limitation associated with these high TP levels, possibly due to sediment resuspension. The "flattening" of the log{dollar}sb{lcub}10{rcub}{dollar} Stage-log{dollar}sb{lcub}10{rcub}{dollar} chlorophyll a relationships as stage increases indicates that the linear form of the relationship does not continue for the entire range of lake stages observed.; Nonparametric regression with Bayesian variable selection and spline interactions provides a flexible framework for addressing the problems identified with the previous work. The use of regression splines allows nonlinear effects to be manifest, while their extension allows inclusion of interactions for which the mathematical form cannot be specified a priori. The use of Bayesian model averaging deals with the problem of model uncertainty that was not addressed in the previous models. The predictive performances of the three models are compared, using set aside data that were not used to specify them. A mechanistic model for prediction of chlorophyll a in Lake Okeechobee under development by the South Florida Water Management District (WASP) is compared and contrasted with the models developed here, though a formal assessment of predictive performance of the WASP model is beyond the scope of this work.