Phytoplankton pigment concentrations retrieved from ocean color satellite observations provide key information on marine ecosystem composition and biogeochemical cycling. Although Machine Learning (ML) approaches have been applied to multi-spectral sensors such as the Moderate Resolution Imaging Spectroradiometer (MODIS) and the Sea-Viewing Wide Field-of-View Sensor (SeaWiFS), no study has systematically exploited the specific band configuration of the Ocean and Land Colour Instrument (OLCI) onboard Copernicus Sentinel-3 for the simultaneous retrieval of multiple diagnostic pigments. This study presents the first systematic benchmark of five ML architectures: Random Forest (RF), eXtreme Gradient Boosting (XGB), Dense Neural Network (DNN), one-dimensional Convolutional Neural Network (CNN), and Bidirectional Long Short-Term Memory Network (BiLSTM) for retrieving 13 pigments of interest to the ocean color community from OLCI-equivalent multi-spectral radiometry. Training and evaluation rely on 185 co-located in situ remote-sensing reflectance and High Performance Liquid Chromatography (HPLC) pigment measurements collected across optically diverse European waters (Mediterranean Sea, Black Sea, Atlantic, and English Channel) within the BiOMaP programme. A novel two-phase Deep Learning (DL) training strategy is introduced, in which pigment-specific modules are first trained independently and subsequently concatenated into a joint multi-output model, preventing dominant, easily retrieved pigments from degrading performance for rarer accessory ones. To extend the methodology to operational satellite products, a transfer learning strategy based on fine-tuning is applied and explicitly benchmarked against training from scratch on a global matchup dataset. Additionally, a spectrally reduced version (5 central wavelengths) is developed and validated for compatibility with the Copernicus-GlobColour and European Space Agency’s Ocean Colour Climate Change Initiative datasets, enabling application of the multi-pigment retrieval framework to a 26-year continuous satellite record. Results indicate that CNN consistently outperforms all other architectures across all evaluation metrics, achieving coefficients of determination of 0.70–0.93 and mean absolute percentage errors of 25–55% across the 13 target pigments, close to the approximate 20% average percent differences of published inter-laboratory HPLC measurement round-robin experiments. The fine-tuned model systematically outperforms the from-scratch version across all 13 pigments, with mean absolute percentage errors generally lower than 70% and approximately 20-25% smaller than those obtained when training from scratch, and a coefficient of determination (R2) decreasing by approximately 0.2 relative to the in situ baseline versus a decrease of 0.3 for the from-scratch model. The spectrally reduced model retains comparable in situ performance and similar transfer learning advantages after fine-tuning. These results provide a quantitative demonstration of the value of high-quality regional in situ radiometry as a pre-training foundation for global satellite pigment retrieval. Together, these findings demonstrate the strong potential of OLCI-based CNN models for operational multi-pigment retrieval across current and future satellite missions and for long-term climatological analysis.