33,640 ocean samples: Machine learning reads water color much more than chlorophyll

By exploiting all multispectral reflectance rather than chlorophyll-a alone, models estimate the pigments that make it possible to identify phytoplankton groups.

Chlorophyll-a concentration is the most widely available measure in ocean satellite observation, but it says little about the taxonomic composition of phytoplankton. The accessory pigments, on the other hand, make it possible to distinguish large groups — still have to be extracted from a limited spectral resolution, which has long been out of reach.

The team led by David Moffat matched 33,640 high-performance liquid chromatography measurements from a global game with reflectance data from ESA's Ocean Colour Climate Change initiative. Two families of models, Random Forest and TabPFN, were trained on complete multispectral reflectance, with a time stratified validation protocol to avoid overestimating performance.

Multispectral models systematically advance approaches that rely only on chlorophyll-a derived from the satellite. However, the gains vary greatly depending on the pigment considered: they are substantial for some, marginal for those who remain strongly correlated with chlorophyll-a and therefore do not provide independent information.