Assessing Microbiological Water Quality Dynamics in Hurricane Florence Floodwaters as a Function of Floodwater Contact Time

Abstract:

Floods connect surface waters with potential pollutant sources, creating risks to public health. Yet, the magnitude of floodwater contamination risk remains poorly understood, particularly in floods of agricultural landscapes. To better understand water quality dynamics in floodwaters of agriculturally-dominated basins, we investigated microbiological water quality conditions in floodwaters across eastern North Carolina (NC) in the wake of Hurricane Florence. Our primary objectives were to (1) estimate floodwater contact time across land use and land cover (LULC) classes, and (2) relate LULC floodwater contact times to observed water quality conditions. We applied Random Forest, a machine learning algorithm, to produce a flood model with temporal resolution of 1 day and spatial resolution of 10 meters to recreate the spatial flood timeline associated with Hurricane Florence. To train the Random Forest model, high-resolution remotely sensed imagery from Planet Labs was used to extract water features through a Normalized Difference Water Index approach and then resampled to 10 meters. An initial set of 15 geophysical (e.g. LULC classes and elevation) and socio-environmental (e.g. social vulnerability and population density) variables were screened as predictors during model building. Specifically, data were used from flood-impacted watersheds in the NC Coastal Plain after Hurricanes Matthew (2016) and Florence (2018), which caused record flooding. Outputs from this model were then used to determine the duration of time that floodwaters made contact with different LULC classes. Finally, LULC floodwater contact times were compared to E. coli measurements made across two sampling events that occurred within one month of Hurricane Florence making landfall. Preliminary findings from this work show that flood hazard and flood frequency did not serve as effective predictors of flooding in the model, while rainfall and elevation data were the most predictive. An investigation into flood hazard and flood frequency data showed that these datasets are likely not detailed and comprehensive enough for effective flood prediction. Ongoing work focuses on evaluating connections between floodwater contact time across LULC classes and microbiological water quality outcomes.