USING MULTIPLE LOGISTIC REGRESSION AND ARTIFICIAL NEURAL NETWORK MODELS TO PREDICT ELEVATED BACTERIA CONCENCTRATIONS IN NATURAL WATERS

This project’s primary focus is to predict “safe for contact” or “unsafe for contact” levels of fecal coliform (FC) and/or Escherichia coli (E. coli) concentrations in natural waters using multiple logistic regression (MLR) and the artificial neural network (ANN) models. Two separate data sets were analyzed and modeled using physiochemical parameters to determine prediction of bacterial concentrations. The first data set from a karst aquifer in the Central Bluegrass Region of Kentucky and the second being the Concho River system in the arid city limits of San Angelo, Texas. Physiochemical parameters used in these models consisted of pH, electrical conductivity, water temperature and precipitation amount within 24 hours prior to sampling; while microbial parameters included E. coli and FC (concentrations in colony forming units (cfu)/100 mL. The level of “unsafe for contact” bacterial counts were determined by applying the Environmental Protection Agency (EPA) guidelines for primary contact standards for both E. coli and FC, which were set as binary dependent variables in the models (microbial parameter for MLR and ANN FC values “unsafe for contact” were set to 1 [i.e., > 200 cfu/100 mL] and values “safe for contact” were set as 0 [i.e., < 200 cfu/100 mL]) with the remainder of parameters considered as independent variables. Binary dependent variable were E. coli values “unsafe for contact” were set to 1 [i.e.,> 126cfu/100 mL] and values “safe for contact” were set as 0 [i.e.,< 126cfu/100 mL]). An MLR model using only physiochemical parameters correctly predicted “safe for contact” conditions 65.6% of the time and “unsafe for contact” conditions 69.2% of the time within the karst system. In the Concho River system, the MLR model correctly predicted “safe for contact levels” 72.4% of the time and “unsafe for contact” 53.9% of the time. Using the ANN model, conditions for “safe for contact” and “unsafe for contact” levels were accurately predicted 66% and 85% of the time respectively in the karst system and 88% and 95% in the river system. The ANN model predicted bacterial contamination more accurately than did the MLR model within both the karst and river system. It is hoped that these models will produce a long-term water quality evaluation tool allowing for determination of general water quality quickly and efficiently.