Towards improved harmful algal bloom forecasts: A comparison of symbolic regression with DoME and stream learning performance

Abstract

Diarrhetic Shellfish Poisoning (DSP) is a global health issue caused by shellfish contaminated with toxins from dinoflagellates, posing significant risks to public health and the shellfish industry. Harmful Algal Blooms (HABs), driven by toxin-producing algae like DSP, require effective monitoring and forecasting systems. Predicting HABs is challenging due to the time-series nature of the problem, influenced by historical seasonal patterns and recent anomalies from meteorological and oceanographic changes. Stream Learning shows promise for handling time-series problems with concept drifts but has yet to be validated for HAB prediction compared to Batch Learning. Limited historical data availability in oceanography highlights the importance of advanced tools like the CROCO ocean hydrodynamic model, which provides high-resolution temporal and spatial data. This study developed a machine learning workflow to predict toxic dinoflagellate (Dinophysis acuminata) cell counts, comparing seven algorithms across two learning paradigms. The CROCO model data addressed historical data gaps. The DoME model, with an average $R^2$ of 0.77 for 3-day-ahead predictions, proved the most effective and interpretable, underscoring the value of model explainability and rigorous comparison methodologies.