Evaluating the added value of physical information for deep-learning estimates of global phytoplankton size classes from satellite

Phytoplankton size structure shapes marine food webs and carbon export, yet satellite ocean-colour retrievals remain subject to substantial uncertainties, since spectrally similar reflectance signals can arise from distinct communities. Here, we assess whether satellite-derived physical variables can complement remote-sensing reflectance (Rrs) to improve global retrievals of total chlorophyll a (total-Chla) and of the micro-, nano-, and pico-Chla size classes. Using a global high-performance liquid chromatography (HPLC) dataset, we trained multilayer perceptrons (MLP) within a controlled framework that combines a spatio-temporal data split to limit leakage, explicit predictor selection without geographical coordinates, and delta-method uncertainty estimates. Compared with the operational OC-CCI product, a MLP using Rrs alone increased R2 by 0.04 for total-Chla, 0.09 for nano-Chla, and 0.33 for pico-Chla, all exceeding the evaluation uncertainty. Among the physical predictors tested, only sea surface temperature (SST) and absolute dynamic topography (ADT) provided consistent additional information; including them further increased R2 for pico-Chla by 0.15. Physical predictors mainly modified biomass partitioning among size classes rather than total-Chla, with the largest effects at high latitudes, where they reproduced the shift from pico- to larger phytoplankton observed by independent flow cytometry in the Southern Ocean. In warm equatorial waters, by contrast, they tended to overestimate pico-Chla. These results show that physical context helps resolve community shifts that are weakly expressed in the optical signal.