Skip to main content
2023 ESA Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.
Back

Contributed Poster

Session: Modeling

PS 32-192 - A multi-model ensemble to explore the limits of phytoplankton bloom prediction

Tuesday, August 8, 2023
5:00 PM – 6:30 PM PDT
Location: ESA Exhibit Hall
Abstract: Model predictions are critical to inform management of natural resources and ensure future provision of ecosystem services. In particular, multi-model ensembles, comprising many different predictive models with varying structures, have proven utility in applications such as fisheries management and predicting species distributions. Multi-model ensemble predictions are often more skillful than predictions from any individual model as models are balanced against each other, such that the ensemble reflects consolidated confidence when models agree and permits hedging of bets when models disagree. Moreover, valuable ecological insight can be gained by comparing predictive performance among models within an ensemble under different environmental conditions, as each model represents a different hypothesis about the ecological drivers and processes related to the focal variable. Despite the value of multi-model ensembles, they have been rarely applied in predictions of freshwater phytoplankton dynamics, which are critically needed for water resource management due to the water quality degradation associated with phytoplankton blooms. Much effort has been dedicated to prediction of phytoplankton dynamics in freshwaters, but predictions are usually generated from only a single model and the predictive skill of most phytoplankton models remains low, especially during bloom events. Here, we use an ensemble of contrasting models to predict phytoplankton biomass in a small drinking water reservoir that exhibits episodic blooms of harmful phytoplankton taxa. We calibrated eight models, including process-based, statistical, and machine learning models, using three years of high-frequency phytoplankton and environmental data. We then assessed model performance over one additional year under different environmental conditions (e.g., in a thermally stratifed vs. mixed water column). We found that ensemble predictions outperformed individual model predictions. Within the ensemble, machine learning and statistical models trained across a range of observed phytoplankton biomass values performed better during times of low biomass, while process-based models and machine learning models trained to optimize fit during times of high biomass performed better during blooms. Our results indicate that different models may be better suited to guide water quality management at different times of year based on environmental conditions such as water temperature, dissolved nitrogen and phosphorus concentrations, and irradiance in surface waters. Altogether, our work can help guide forecasting efforts by applying multi-model ensembles to optimize natural resource management.