This benchmark dataset contains the physical data used as predictors to reconstruct global chlorophyll-a concentrations (Chl, a proxy of phytoplankton biomass) in Roussillon et al., as well as the reference satellite Chl target fields. The nine physical predictors' data (Short-Wave radiations, Sea Surface Temperature, Sea Level Anomaly, Zonal and meridional surface currents, Zonal and meridional surface wind stress, Bathymetry, Binary continental mask) were extracted from publicly available datasets over [1998-2015] and resampled to the same spatio-temporel resolution as Chl, i.e. monthly on a 1°x1° grid between 50°N and 50°S. Missing values were gap-filled using the heat diffusion equation. Each variable was normalized by substracting its mean from the original values and dividing by its standard deviation over [1998-2015]. This dataset was used to train and validate the Multi-Mode Convolutional Neural network (CNNMM8) introduced in Roussillon et al. ; reconstructed monthly Chl fields over the [2012-2015] test period are also provided here. We hope this benchmark dataset can help to promote the improvements of methods as well as the emergence of new ideas, as building datasets is sometimes more time-consuming than the implementation of machine learning tools themselves. This would also facilitate the quantitative comparison of models performances' on the exact same datasets.

The presented database includes 0-1000 m vertical profiles of bio-optical and biogeochemical variables acquired by autonomous profiling Biogeochemical-Argo (BGC-Argo) floats. Data have been collected between October 2012 and January 2016, around local noon, in several oceanic areas encompassing the diversity of ocean’s trophic environments. The database includes profiles of downward irradiance at 3 wavelengths (380, 412 and 490 nm), photosynthetically available radiation, chlorophyll a concentration, fluorescent dissolved organic matter, and particle light backscattering at 700 nm. All variables have been quality controlled following specifically-developed procedures, that aimed to support biogeochemical and bio-optical applications at the global scale. Data corruption by biofouling and any instrumental drift has also been verified. Moreover, to allow users for different biogeochemical applications, vertical profiles of chlorophyll a and particle light backscattering at 700 nm have been presented before and after advanced processing (e.g., non-photochemical quenching correction, identification of spikes). Vertical profiles of temperature and salinity associated to these bio-optical data are also provided although they have been only quality-controlled for sensor issues related to bio-fouling and instrumental drift.

The database here presented includes quality-controlled bio-optical and biogeochemical variables derived from autonomous profiling Biogeochemical Argo float measurements, and supports field and remote ocean color applications. Data represent the first optical depth (i.e., the layer of interest for satellite remote sensing) and have been collected between October 2012 and January 2016, around local noon, in several oceanic areas encompassing the diversity of ocean’s trophic environments. In addition to the first optical depth (as derived from the in situ measured depth of the euphotic layer), the presented database is compiled with data of diffuse light attenuation coefficients of photosynthetically available radiation and downward irradiance at 3 wavelengths (380, 412 and 490 nm), chlorophyll a concentration, fluorescent dissolved organic matter, and particle light backscattering at 700 nm. Associated errors to each variable are also reported.