Deep Learning for Enhanced Ocean Color Remote Sensing: A Foundation Model Approach

Ocean color remote sensing has advanced over the past 40 years, with algorithms primarily rooted in radiative transfer theory to estimate various ocean color properties. To address the diversity and region-specific distribution of these properties, numerous algorithms have been developed; however, each is tailored to specific variables. In recent years, deep learning has demonstrated remarkable progress in this field, but a general foundational model for robust ocean color product generation remains lacking. To address this limitation, this study introduces the ocean color deep-learning foundation model (OCFM), a comprehensive framework designed to extract multiple global ocean color properties from satellite observations. The development of OCFM follows three sequential phases: pretraining with operational satellite products to learn fundamental theoretical relationships, fine-tuning with in situ measurements to align with the real ocean state, and deploying the trained OCFM for flexible retrieval of ocean color properties. Quantitative evaluation shows that, in end-user application tests, the model achieved a coefficient of determination ( ${R} ^{\mathbf {2}}$ ) of 0.90 for primary productivity and 0.71 for water clarity, with corresponding mean absolute percentage differences of 44.08% and 17.28%, respectively. OCFM provides a more efficient solution for few-shot and unevenly distributed samples compared with traditional retrieval algorithms. Even with limited user resources and small sample sizes, additional downstream training with the trained OCFM can achieve state-of-the-art performance in retrieving ocean color properties. This study highlights the potential of a deep-learning foundation model for generalizable and few-shot ocean color retrieval.