How can convolutional neural networks help account for the impact of mesoscale ocean structures on phytoplankton distributions ?

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—processes central to the so-called 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.

In many places of the ocean, climate change is expected to result in warmer and more oligotrophic surface waters. This should influence the composition of phytoplankton communities, displacing the dominance towards smaller-sized organisms. This community change would, in turn, affect the services that phytoplankton provides to humans, such as carbon sequestration and the support of fisheries. This is why large-scale monitoring of the abundance and composition of phytoplankton communities is essential for assessing ocean health.

Ocean color sensors on satellites can provide such large-scale monitoring. Current products comprise daily, 4 km-resolution maps of chlorophyll-a concentration (the most widely used estimator of phytoplankton biomass) and its distribution across three phytoplankton size classes (pico-, nano- and micro-phytoplankton). The algorithms to produce these maps rely on the relationships between in situ measurements and reflectances at a few wavelengths, over a few pixels, matching the time and location of the in situ measurement. While incredibly useful, they still yield 30% error for total chlorophyll-a concentration and 40% for community composition. 

Notably, they cannot capture mesoscale oceanic structures, such as fronts and eddies, while they significantly influence phytoplankton production and distribution. The signature of these structures can be observed through infrared and radio wave satellite data and spans tens to hundreds of kilometers. 

In this work, we use deep learning models to (1) naturally combine reflectances with other satellite-derived variables that describe ocean physics (sea surface temperature, sea level anomalies, etc.), (2) use arrays of data covering dozens of kilometers around the in situ observations, that can be summarized through convolutional layers, instead of a single point as input. These two approaches should enable us to capture the effect of mesoscale oceanic structures on the abundance and composition of phytoplankton. The models developed already result in  more accurate and nuanced products that can offer valuable information for policymakers and stakeholders, particularly in fisheries management, where understanding plankton distribution is key to assessing fish stock health and ecosystem resilience in a changing climate.