Phytoplankton plays a key role in the carbon cycle and supports the oceanic food web. While its seasonal and interannual cycles are rather well characterized owing to the modern satellite ocean color era, its longer time variability remains largely unknown due to the short time-period covered by observations on a global scale. With the aim of reconstructing this longer-term phytoplankton variability, a support vector regression (SVR) approach was recently considered to derive surface Chlorophyll-a concentration (Chl, a proxy of phytoplankton biomass) from physical oceanic model outputs and atmospheric reanalysis. However, those early efforts relied on one particular algorithm, putting aside the question of whether different algorithms may have specific behaviors. Here, we show that this approach can also be applied on satellite observations and can even be further improved by testing performances of different machine learning algorithms, the SVR and a neural network with dense layers (a multi-layer perceptron, MLP). The MLP, thanks to its ability to capture complex non-linear relationships, outperforms the SVR to capture satellite Chl spatial patterns (correlation of 0.75 vs. 0.65 on a global scale, respectively) along with its interannual variability and trend, despite an underestimated amplitude. Among deep learning algorithms, neural network such as MLP models appear to be promising tools to investigate phytoplankton long-term time-series.

The authors wish to make the following corrections to the paper

Phytoplankton plays a key role in the carbon cycle and constitutes the basis of the marine food web. Its seasonal and interannual cycles are relatively well-known on a global scale thanks to continuous ocean color satellite observations acquired since 1997. The satellite-derived chlorophyll-a concentrations (Chl-a, a proxy of phytoplankton biomass) time series are still too short to investigate phytoplankton biomass low-frequency variability. However, it is a vital prerequisite before being able to confidently detect anthropogenic signals, as natural decadal variability can accentuate, weaken or even mask out any anthropogenic trends. Machine learning appears as a promising tool to reconstruct Chl-a past signals (including periods before satellite Chl-a era), and deep learning models seem particularly relevant to explore the spatial and/or temporal structure of the data.Here, different neural network architectures have been tested on a 18-year satellite and re-analysis dataset to infer Chl-a from physical predictors. Their ability to reconstruct spatial and temporal (seasonal and interannual) variations on a global scale will be presented. Convolutional neural networks (CNN) better capture Chl-a spatial fields than models that do not account for the structure of the data, such as multi-layer perceptrons (MLPs). We also assess how the selection of training period may affect the reconstruction performance. This is a necessary step before being able to reconstruct any past Chl-a multi-decadal time series with confidence, which is the ultimate goal of this work.Our study also addresses the carbon footprint associated with the use of GPU resources when training the CNN. GPUs are energy intensive, and their use in geosciences is expected to grow fast. Systematically reporting the computational energy costs in the geoscience community studies would provide an overview of models energy-efficiency on different kinds of datasets and may encourage actions to reduce consumption when possible.

Phytoplankton plays a key role in the carbon cycle and fuels marine food webs. Its seasonal and interannual variations are relatively well-known at global scale thanks to satellite ocean color observations that have been continuously acquired since 1997. However, the satellite-derived chlorophyll-a concentrations (Chl-a, a proxy of phytoplankton biomass) time series are still too short to investigate phytoplankton biomass low-frequency variability. Machine learning models such as support vector regression (SVR) or multi-layer perceptron (MLP) have recently proven to be an alternative approach to mechanistic ones to reconstruct Chl-a past signals (including periods before the satellite era) from physical predictors, but they remain unsatisfactory. In particular, the relationships between phytoplankton and its physical surrounding environment are not homogeneous in space, and training such models over the entire globe does not allow them to capture these regional specificities. Moreover, if the global ocean is commonly partitioned into biogeochemical provinces into which phytoplankton growth is supposed to be governed by similar processes, their time-evolving nature makes it difficult to impose a priori spatial constraints to restrict the learning phase on specific areas. Here, we propose to overcome this limitation by introducing spatial multi-modalities into a convolutional neural network (CNN). The latter can learn with no particular supervision several spatially weighted modes of variability. Each of them is associated with a CNN submodel trained in parallel, standing for a mode-specific response of phytoplankton biomass to the physical forcing. Beyond improving performance reconstruction, we will show that the learned spatial modes appear physically consistent and may help to get new insights into physical-biogeochemical processes controlling phytoplankton repartition at global scale.