Comparative Analysis of Machine Learning Methods for Modeling Sargassum Prevalence in the Gulf of Mexico

Keywords: Sargassum, Machine Learning, Random Forest, Isolated Random Forest, ErisNet, Prithvi, Gulf of Mexico, Environmental Modeling
Abstract Type: Paper Abstract

Abstract

Rafts of pelagic Sargassum have been washing ashore in the Gulf of Mexico and Caribbean Sea with unprecedented frequency and abundance over recent years. Massive beaching events deposit volumes of Sargassum along shorelines. On land, Sargassum smothers vulnerable ecosystems like mangrove swamps, prohibits sea-turtle access to nesting sites, and produces hypoxic conditions, simultaneously disrupting local fishing industries and endangering fish populations and other marine life. Sargassum releases hydrogen sulfide gas and ammonia as it decays, causing significant health concerns for local inhabitants and triggering economic upheaval for coastal communities whose livelihoods depend on tourism. In this study, we compare the efficacy of four supervised learning methods, Random Forest, Isolated Random Forest, ErisNet, and Prithvi, for modeling Sargassum prevalence in Miami Beach and Cancun Beach over a span of four months. We employ each of the machine learning techniques using Planet’s Dove and SuperDove observations, selected for high spatial resolution (3 meter) and daily revisit-frequency, to identify and extract locations and quantities of Sargassum both in near-shore water and on land for each area of interest. The results from this study include visualization, statistical analysis, and discussion of the results from each machine learning method. Statistical analysis of the performance of the four machine learning methods includes comparative evaluation metrics for accuracy, precision, recall, and F1-scores. We discuss the performance of the different models to assess suitability for Sargassum monitoring with respect to near-shore or shallow waters, beached Sargassum, and floating Sargassum.