Exploratory Machine Learning Modeling of Cyanobacteria and Cyanotoxin Transport in the Raritan Basin of New Jersey

Harmful algal blooms (HABs), often dominated by toxin-producing cyanobacteria, pose increasing ecological, economic, and public health risks in freshwater systems. Microcystins, among the most common cyanotoxins, contaminate drinking water sources and threaten aquatic life, making accurate prediction of bloom dynamics essential for early warning and mitigation. This study examines three key biological targets in the Raritan Basin Water Supply Complex (RBWSC) of New Jersey: microcystin concentration (μg/L), total cyanobacteria abundance (cells/mL), and total phytoplankton biomass (cells/mL). Field measurements collected by the U.S. Geological Survey from August 2020 to August 2021 included biological, chemical, and physical water-quality parameters at nine RBWSC monitoring sites. We evaluate statistical, deep learning, and ensemble approaches for concentration regression and find that all methods perform inadequately under temporal validation due to sparse sampling and skewed target ranges. We therefore reframe the task as binary regulatory threshold exceedance aligned with the U.S. Environmental Protection Agency (EPA) and the World Health Organization (WHO) water-quality guidelines to support intervention decisions. Binary classification achieves F1 = 0.900 and Matthews Correlation Coefficient MCC = 0.883 for phytoplankton and F1 = 0.889 (MCC = 0.873) for cyanobacteria, with constrained-depth tree ensembles achieving F1 = 0.800 (MCC = 0.808) for microcystin detection, all exceeding the 70-85% classification accuracy seen in HAB prediction literature. Extending to two-week forecasting, gradient-boosted trees incorporating prior blooms achieve F1 = 0.923 for cyanobacteria, surpassing the same-time classification performance and confirming the temporal persistence of bloom conditions. Leave-one-site-out cross-validation confirms spatial generalizability across unseen monitoring stations with F1 : 0.65-0.75. Framing water safety as a binary classification task based on regulatory thresholds allows for the development of a reliable HAB early warning system that can be used to inform water treatment and public health response.