Machine learning (ML) models are capable of providing system-specific information based on observed patterns in real-world datasets. In this study, we utilize random forest (RF), a machine learning algorithm that constructs an ensemble of multiple decision trees to make predictions, for the case study of Heterosigma akashiwo blooms in the Hudson-Raritan Estuary (HRE). Environmental data for the model, from six sites within the estuary from (Rothenberger & Calomeni, 2010) to 2022, were collected within an adaptive monitoring framework that emphasizes resource management decision-making. The RF model had an out-of-bag prediction error rate of about 11.83% with lower error for low-density months (7.23%) than high-density months (36.26%). While these error rates are comparable to similar studies and indicate good performance for predicting rare events in complex systems, further improvements to the model may enhance prediction. RF model results identified three weather-related factors (i.e., river discharge and precipitation) and two biotic factors (i.e., abundance of Heterocapsa rotundata and Chlamydomonas spp.) as the most important predictors of H. akashiwo abundance in the HRE. We then paired results of the RF model with a time series analysis of H. akashiwo abundance and potentially high-leverage environmental predictors. Significant declines in dissolved oxygen and pH and significant increases in ammonium in the system suggest continued water quality degradation that will likely continue to favor harmful algal blooming species. Successful execution of this type of analysis suggests that the practical, predictive ML approach can be extended beyond the HRE to other ecosystems of interest with similar data structures.
