Constructing predictive water quality models for fecal coliforms in shellfish growing waters

Background/Question/Methods

Coastal shellfish growing waters are regularly tested for fecal coliform bacteria to maintain public health and safety standards for shellfish consumers. In Florida, the Department of Agriculture and Consumer Services (FDACS) regulates shellfish growing waters through closures of harvesting areas after a potential pollution event has occurred. FDACS has maintained a digitized water quality database going back to 1978 containing the results of fecal coliform water sampling. While this data is used for day-to-day management, the quantity and resolution of data provides the opportunity for prediction of fecal coliform concentrations in Florida’s shellfish growing areas. The data consists of 10 years of water samples from 2,247 FDACS sampling stations within 40 harvesting areas that are grouped into 5 different jurisdictions, which we partitioned into a training and test data set. Using machine learning approaches, we (1) identified key drivers of fecal coliform dynamics between harvesting areas, (2) estimated fecal coliform concentrations, and (3) predicted shellfish harvesting area closures.

Results/Conclusions

We trained Random Forest regression models for each of the 5 harvesting area jurisdictions to predict mean fecal coliform concentrations across sampling stations. We then used our predicted fecal coliform concentrations to classify areas as either closed or open based on the National Shellfish Sanitation Program guideline threshold. Predictors included rainfall, wind speed and direction, length of natural and modified waterways, land use and land cover, soil drainage, tidal stage, air temperature, river stage, and season. We identified important features contributing to fecal coliform measurements with variable importance scores, which are calculated during the Random Forest model training process. These predictors of importance provide insight into the locally specific dynamics that drive fecal coliform loads in the coastal areas. We used 10-fold cross-validation to train the models and evaluated model performance using Cohen’s Kappa coefficient. Ultimately this study informs future management strategies, provides a programmatic and statistical workflow for the development of similar aquaculture management support tools, and reveals new insights into the complexities of estuarine and coastal systems.