A Multi-Mode Convolutional Neural Network to reconstruct satellite-derived chlorophyll-a time series in the global ocean from physical drivers

Time series of satellite-derived chlorophyll-a concentration (Chl, a proxy of phytoplankton biomass), continuously generated since 1997, are still too short to investigate the low-frequency variability of phytoplankton biomass (e.g. decadal 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 synoptic past time-series before the satellite era from physical predictors. Nevertheless, the relationships between phytoplankton and its physical surrounding environment were implicitly considered homogeneous in space, and training such models on a global scale does not allow one to consider known regional mechanisms. Indeed, the global ocean is commonly partitioned into biogeochemical provinces (BGCPs) into which phytoplankton growth is supposed to be governed by regionally-”homogeneous” processes. The time-evolving nature of those provinces prevents imposing a priori spatially-fixed boundary constraints to restrict the learning phase. Here, we propose to use a multi-mode Convolutional Neural Network (CNN), which can spatially learn and combine different modes, to globally account for interregional variabilities. Each mode is associated with a CNN submodel, standing for a mode-specific response of phytoplankton biomass to the physical forcing. Beyond improving performance reconstruction, we show that the different modes appear regionally consistent with the ocean dynamics and that they may help to get new insights into physical-biogeochemical processes controlling phytoplankton spatio-temporal variability at global scale.