Predicting water quality at Santa Monica Beach: Evaluation of five different models for public notification of unsafe swimming conditions

Bathing beaches are monitored for fecal indicator bacteria (FIB) to protect swimmers from unsafe conditions. However, FIB assays take ∼24 h and water quality conditions can change dramatically in that time, so unsafe conditions cannot presently be identified in a timely manner. Statistical, data-driven predictive models use information on environmental conditions (i.e., rainfall, turbidity) to provide nowcasts of FIB concentrations. Their ability to predict real time FIB concentrations can make them more accurate at identifying unsafe conditions than the current method of using day or older FIB measurements. Predictive models are used in the Great Lakes, Hong Kong, and Scotland for beach management, but they are presently not used in California - the location of some of the world's most popular beaches. California beaches are unique as point source pollution has generally been mitigated, the summer bathing season receives little to no rainfall, and in situ measurements of turbidity and salinity are not readily available. These characteristics may make modeling FIB difficult, as many current FIB models rely heavily on rainfall or salinity. The current study investigates the potential for FIB models to predict water quality at a quintessential California Beach: Santa Monica Beach. This study compares the performance of five predictive models, multiple linear regression model, binary logistic regression model, partial least square regression model, artificial neural network, and classification tree, to predict concentrations of summertime fecal coliform and enterococci concentrations. Past measurements of bacterial concentration, storm drain condition, and tide level are found to be critical factors in the predictive models. The models perform better than the current beach management method. The classification tree models perform the best; for example they correctly predict 42% of beach postings due to fecal coliform exceedances during model validation, as compared to 28% by the current method. Artificial neural network is the second best model which minimizes the number of incorrect beach postings. The binary logistic regression model also gives promising results, comparable to classification tree, by adjusting the posting decision thresholds to maximize correct beach postings. This study indicates that predictive models hold promise as a beach management tool at Santa Monica Beach. However, there are opportunities to further refine predictive models.