Predicting estuarine fecal indicator bacteria concentrations using sensor data

Coastal surface water quality has direct impacts on human health through recreational use and shellfish sanitation, to prevent exposure to contaminated waters, monitoring fecal indicator bacteria (FIB) is imperative. However, current monitoring is resource-intensive, leading to coarse data and substantial lag times between sample collection and reporting. As sensor technology improves, our ability to monitor water quality using optical sensors expands, presenting the opportunity to predict FIB at high temporal frequencies using in-situ multi-parameter sonde data. Yet, research is needed to determine if sensor data can reliably predict FIB concentrations. Our study objectives were to (1) collect FIB data over dynamic conditions alongside high-frequency sonde measurements, (2) develop and train statistical models to predict FIB levels using sonde data, and (3) identify optimal covariates for predicting FIB concentrations. We deployed an EXO2 sonde on the estuarine Newport River in Beaufort, NC that recorded measurements every 15 minutes for tryptophan-like fluorescence (TLF), salinity, temperature, dissolved oxygen, fDOM, turbidity, total algae, and pH for three weeks in summer of 2021. We collected grab samples daily and performed 4 intensive samplings to capture baseline and storm-flow conditions in high resolution. We quantified Enterococci and E. coli in these samples. From the sonde and conventional FIB data, we created a set of statistical and machine learning models to estimate FIB continuously and categorically (i.e., above/below threshold). Models developed for Enterococci have higher performance statistics, potentially due to a larger sample size. Of the continuous models created, a random forest model for Enterococci using all covariates as predictors had the best performance (training: R2 = 0.65; testing R2 = 0.67). An Enterococci logistic regression model using each sonde covariate as a predictor had a testing sensitivity of 4/6 and specificity of 19/20. For the Enterococci models, TLF, salinity, temperature and FDOM had high predictor importance for both regression and random forest models, indicating a reduced number of predictors could be used in the models. Overall, our results show a sonde-based nowcasting system for estuarine FIB is feasible, but further research is needed.