Extending the inputs of deep learning models to capture the mesoscale context and better predict phytoplankton community composition

Phytoplankton is a central component of marine ecosystems. It contributes to biogeochemical cycles by absorbing carbon through photosynthesis at the ocean surface and transporting it deeper through sinking and subduction—hence contributing to the biological carbon pump. Plankton also represents the first link in marine food webs, supporting a wide range of marine life, from other plankters to the most productive fisheries on earth.

Satellites can help monitor phytoplankton over large-scales thanks to ocean color sensors. Current products provide daily, 4 km-resolution fields of chlorophyll-a concentration (Chla, the most widely used estimator of phytoplankton biomass) as well as its distribution in a few groups, hence estimating broad community composition. To produce these operational maps, the concentration of pigments measured by HPLC (High-Performance Liquid Chromatography) from in situ samples is regressed on reflectances at a few wavelengths matched to those samples in space and time. While incredibly useful, these models still display 30% error for Chla and at least as much when predicting community composition.

Numerous studies have shown the importance of considering mesoscale ocean structures, such as fronts and eddies, as they have a significant influence on the production and distribution of phytoplankton. These structures span tens to hundreds of kilometers and can be observed through ocean color but also infrared and radio wave satellite data.

In this work, we develop a deep learning model to predict the concentration of three phytoplankton size classes: pico-, nano-, and micro-phytoplankton. The in situ values are derived from over 7000 HPLC measurements spanning the globe, from 1997 to 2021. We use a Multi-Layer Perceptron to naturally combine reflectances with other satellite-derived variables that describe ocean physics (sea surface temperature, sea level anomalies, etc.) as input. The MLP is preceded by convolutional layers to summarise arrays of the input variables covering dozens of kilometers around the in situ observations. These two approaches are meant to capture the effect of mesoscale oceanic structures on the abundance and composition of phytoplankton.

This approach improves the estimation of phytoplankton communities on a global scale. It paves the way for in-depth studies on the influence of mesoscale structure in specific oceanic regions. Furthermore, it lays the groundwork for the future integration of the temporal dimension into the model, enabling a more comprehensive representation of ecological dynamics.