ABSTRACT

The Arthur R. Marshall Loxahatchee Wildlife Refuge (Refuge) was established in 1951 through a license agreement between the South Florida Water Management District and the U.S. Fish and Wildlife Service (USFWS) as part of the Migratory Bird Conservation Act. Under the license agreement, the State of Florida owns the land of the Refuge and the USFWS manages the land. Fifty-seven miles of levees and borrow canals surround the Refuge. Water in the canals surrounding the marsh is controlled by inflows and outflows through control structures. The transport of canal water with higher specific conductance and nutrient concentrations to the interior marsh has the potential to alter critical ecosystem functions of the marsh.

Data-mining techniques were applied to 12 years (1995–2006) of historical data to systematically synthesize and analyze the dataset to enhance the understanding of the hydrology and water quality of the Refuge. From the analysis, empirical models, including artificial neural network (ANN) models, were developed to answer critical questions related to the relative effects of controlled releases, precipitation, and meteorological forcing on water levels, specific conductance, and phosphorous concentrations of the interior marsh. Data mining is a powerful tool for converting large databases into information to solve complex problems resulting from large numbers of explanatory variables or poorly understood process physics. For the application of the linear regression and ANN models to the Refuge, data-mining methods were applied to maximize the information content in the raw data. Signal processing techniques used in the data analysis and model development included signal decomposition, digital filtering, time derivatives, time delays, and running averages. Inputs to the empirical models included time series, or signals, of inflows and outflows from the control structures, precipitation, and evapotranspiration. For a complex hydrologic system like the Refuge, the statistical accuracy of the models and predictive capability were good. The water-level models have coefficient of determination (R² values ranging from 0.90 to 0.98. The R² for the specific conductance model is 0.82, and the R² for the total phosphorus model is 0.51. The accuracy of the models was attributable to the quantity and quality of the available data.

To make the models directly available to all stakeholders, an easy-to-use decision support system (DSS) called the Loxahatchee Artificial Neural Network Model (LOXANN) DSS was developed as a spreadsheet application that integrates the historical database, linear regression and ANN models, model controls, streaming graphics, and model output. The LOXANN DSS allows Refuge managers and other users to easily execute the water level, specific conductance, and phosphorous models to evaluate various water-resource management scenarios. The user is able to choose from three options in setting the control-structure flows: as a percentage of historical flow, as a constant flow, or as a user-defined hydrograph. Output from the LOXANN DSS includes tabular time series of predictions of the measured data and predictions of the user-specified conditions. A three-dimensional visualization routine also was developed that displays longitudinal specific conductance conditions.

Two scenarios were simulated with the LOXANN DSS. One scenario increased the historical flows at four control structures by 40 percent. The second scenario used a user-defined hydrograph to set the outflow from the Refuge to the weekly average inflow to the Refuge delayed by 2 days. Both scenarios decreased the potential of canal water intruding into the marsh by decreasing the slope of the water level between the canals and the marsh.

First posted December 30, 2010