The public may be exposed to elevated health risks when using recreational surface waters due to pathogen loadings that originate from overland runoff and combined sewer overflows, especially after a storm event. Hence, water quality is monitored regularly using indicator organisms such as E. coli and fecal coliforms. Since the analysis time required for microbial samples range from 24 to 48 hours, there is a need to develop models that can quickly predict and/or classify the water quality. Previous studies have developed regression and artificial neural network (ANN) models for predicting microbial concentrations that can then be classified based on the pertinent water quality standards, but these models tend to suffer from high false negative rates. The primary objectives of this research are to develop a classification model using learning vector quantization (LVQ), which directly classifies the samples thereby avoiding the prediction step, and to compare the LVQ performance with regression and ANN techniques. The second objective is to identify the more important explanatory variables that provide adequate performance of the algorithms.
The models were developed using data collected at the Larz Anderson Bridge site on the Charles River.
A preliminary analysis was performed to compare the classification efficiencies of the multivariate linear regression (MLR), ANN, and LVQ models using the same three explanatory variables suggested by Eleria and Vogel (2005), which includes the lag-1 fecal coliform concentration as an explanatory variable. All three models performed well in predicting safe conditions (true negative rates > 90%) for both the primary and secondary recreation standards. While the MLR and ANN models had false negative rates between 35% to 55%,the LVQ model produced lower false negative rates between 12% to 18%. The primary drawback of the proposed models is the use of the lag1 fecal coliform concentrations as an explanatory variable, which requires too much time to be adequately included in a predictive model.
Additional research was performed to determine the appropriate explanatory variables to adequately classify the water quality without using the fecal coliform data. The input selection methodology included ranking the explanatory variables (using a different approach for each model) and individually removing the least important input variables. The results provided a tradeoff curve between the model performance and the number of variables included in the model. While all three models were capable of classifying the non-violated samples (> 90%), only the LVQ model had reasonably low false negative rates (< 20%); the MLR and ANN models had false negative rates ranging between 35% to 50%. In addition to evaluating the model performance, the input selection process provided insight into the explanatory variables, which indicated that discharge (from the current or previous day), rainfall during the last week, and the time since the last moderate rainfall were important parameters in predicting and/or classifying microbial water quality. Overall, the LVQ approach appeared to be a suitable solution for the development of a model to classify recreational water quality with a limited number of explanatory variables.