PROJECT TITLE: Modeling Estuarine Salinity Using Artificial Neural Networks 


ABSTRACT: Computer models serve as an important tool for prediction of the response of estuarine ecosystems to anthropogenic and natural changes. This study demonstrated for the first time that artificial neural networks (ANNs) can serve as an alternative modeling tool to simulate estuarine salinity as opposed to traditional physically-based hydrodynamic models. The multilayer perceptron algorithms were programmed (using Java) in an object-oriented manner, allowing for flexibility to easily change the structure of ANNs and to carry out multiple tests in a personal computer.  Using eight years of data collected in the Loxahatchee River, a federal designated Wild and Scenic River in Florida, where saltwater intrusion has been of concern for decades, the ANNs were successfully trained and tested to simulate salinity at three ecologically important locations in the river. The network used freshwater inflow, rainfall, and tide as inputs.  The r2 values ranged from 0.83 to 0.90 during the training period (11/2003-5/2007) and 0.57 to 0.77 in the testing period (6/2007-10/2011).  The trained and tested ANNs were further applied to predict historical salinity from 1972-2011 for evaluation of saltwater intrusion. The result indicated that the spatial and temporal patterns of saltwater intrusion were associated with watershed drainage alterations and regional hydrology.  It further explained the observed increasing mortality of bald cypress and encroachment of mangroves in the tidal floodplain of the river.  Future research can be conducted to compare the versatility of different network structures such as recurrent neural networks or radial basis function networks.

VERSION or DATE: 2013

AUTHORS: COPYRIGHT - CHRISTOPHER WAN