Deep learning applied to ocean color to reconstruct long time-series of phytoplankton biomass in the global ocean

Elodie Claire Martinez1, Anwar Brini1, Thomas Gorgues1, Ronan Fablet2, Lucas Drumetz3, Pierre Tandeo2 and Guillaume Maze4, (1)IRD, Laboratory for Ocean Physics and Satellite remote sensing, Plouzane, France, (2)Telecom Bretagne, Brest, France, (3)IMT Atlantique, Brest, France, (4)IFREMER, Laboratory for Ocean Physics and Satellite remote sensing, France
Phytoplankton plays a key role in the carbon cycle and support the oceanic food web. While their seasonal and interannual cycles are rather well characterized owing to the modern satellite ocean color era, their longer time variability remains largely unknown because of the lack of long-term observations at global scale and uncertainties in numerical biogeochemical simulations. Taking advantage from the strong bottom-up (i.e., physical) forcing on the biological variability in the global ocean, a deep learning approach has been used to reconstruct the satellite derived chlorophyll concentration (Chl, a proxy of phytoplankton biomass) from physical oceanic and atmospheric satellite and reanalysis data over 1998-2015. Several machine learning algorithms, mainly Support Vector machine Regression (SVR), Multi-Layer Perceptron (MLP) and Convolutional Neural Network (CNN), as well as sensitivity tests have been investigated and their skills compared. Correlations between ocean color vs. reconstructed Chl show a better performance of the MLP than the CNN to reconstruct Chl, then the SVR. The ability of the different statistical approaches to capture the Chl seasonal and interannual signal was further investigated through Empirical Orthogonal Function analysis (EOFs). Both MLP and CNN are performant to accurately capture the Chl seasonal signal although the MLP underestimates this mode of variability and inversely for the CNN. On the other hand, the SVR approach fails. The MLP also better succeeds in reproducing the El Nino Southern Oscillation related Chl interannual variability at global scale. Deep learning-based models appear to be an alternative to the conventional physical-biogeochemical models to access and investigate phytoplankton long-term time-series. Further investigations are still needed, especially considering the CNN, prior reconstructing Chl multi-decadal time series.