Deep-Learning Based Seasonal Chlorophyll Forecasting in the Tropical Atlantic

Accurate biogeochemical forecasting of the oceans is crucial for ecosystem and fisheries management. In this study, we propose a methodology for chlorophyll forecasting at a seasonal scale by leveraging physical ocean forecasts and autoencoders, a type of convolutional neural network. Chlorophyll-a is a key indicator of phytoplankton biomass, and it offers the advantage of being relatively easy to measure at a large scale using satellite-based color sensors. Specifically, this approach focuses on estimating surface chlorophyll concentrations in the Tropical Atlantic from forecasted physical properties of the ocean: sea surface temperature and salinity, sea surface height, and mixed layer depth. 

The proposed method is trained on the GlobColour dataset, a cloud-free, merged chlorophyll concentration output from various sensors, from 1998 to 2009. Seven-month forecasts are performed for the period 2010–2020, with monthly initialization. We show that we can skillfully integrate data from the ECMWF’s long-range forecasting system, SEAS5, to predict multiannual and seasonal chlorophyll levels. Evaluation against 2010–2020 GlobColour data demonstrates the autoencoder’s skill in capturing spatial and temporal patterns. Seasonal performance was assessed for regions of interest, including the Equatorial and Senegal-Mauritania upwelling regions, the Inter-Tropical Convergence Zone (ITCZ), and the Northern Atlantic. The neural network consistently outperforms the biogeochemical reanalysis of reference in measured skill and has the additional advantage of being less resource-intensive than traditional models. 

These results further confirm the potential of deep-learning techniques in operational oceanographic applications. Future work will focus on expanding this approach to generating global-scale, multi-nutrient forecasts in the context of the European Digital Ocean Twin (EDITO) project, which aims to provide an environment to exploit this type of machine-learning based simulation algorithms.