Understanding Drivers of Phytoplankton Variability in the Black Sea Using Convolutional Neural Networks

Extreme marine biological events, such as harmful algal blooms and mass mortalities, are increasingly driven by climate variability and anthropogenic pressures, profoundly impacting marine ecosystems. The Black Sea, with its distinct stratification, salinity gradients, and diverse phytoplankton functional groups, is particularly vulnerable to these changes. Understanding and forecasting the interactions between physical, chemical, and biological variables in this region is crucial for effective ecosystem management.

We present a neural network-based surrogate modeling framework to analyze and predict the dynamics of the Black Sea ecosystem. A 3D convolutional encoder-decoder network is trained on simulation data (1950–2014) produced be the University of Liège, including daily basin-scale values of temperature, salinity, nutrients, chlorophyll, and phytoplankton biomass. The model processes time series of spatial maps as input and predicts chlorophyll concentrations and the distributions of phytoplankton functional groups for the subsequent two weeks.

This approach efficiently captures complex interdependencies between variables, offering a computationally efficient alternative to traditional process-based models. By perturbing input variables, the model identifies key drivers of chlorophyll variability, enabling rapid scenario testing to explore the impacts of environmental changes on the ecosystem.

Our findings demonstrate the potential of neural network-based surrogate models to advance understanding of phytoplankton dynamics and support decision-making in marine ecosystem management.