Using Machine Learning to Predict Phytoplankton Blooms in the Salish Sea

Abstract Strong phytoplankton blooms occur every spring in the Salish Sea, but vary significantly, spatially and temporally. Natural variability and climate change make the prediction of the bloom challenging, as they affect both atmospheric drivers and nutrient interactions that contribute to the growth of phytoplankton. Understanding these drivers is essential for a successful bloom prediction. Here, we used a 25‐year data set (18 training, 4 evaluation, and 1 independent year), consisting of 7 atmospheric drivers and 4 phytoplankton variables for the Salish Sea, to build Machine Learning (ML) models that are able to emulate phytoplankton blooms. For phytoplankton productivity, we employed histogram gradient boosting models, whereas for the biomass of phytoplankton, we employed functional regression models, to capture its dynamic nature and dependence on the atmospheric drivers. Feature selection was implemented to identify the optimal set of input features for each model. By using only meteorological drivers (and in some cases spatiotemporal variables) as inputs, the proposed ML models are able to efficiently capture spatial and temporal patterns of phytoplankton blooms over the Salish Sea. Different phytoplankton variables (e.g., biomass, production rates) and different regions require different input features (e.g., precipitation), with shortwave solar radiation being the most frequently used atmospheric driver.