Conference Agenda

Ocean carbon research faces a persistent challenge: high-quality in-situ measurements are extremely sparse and expensive to collect, yet these data are essential for understanding marine primary production and its role in global climate processes. Current satellite-based approaches struggle with limited validation data, a problem recognised by the IPCC as a major constraint on ocean carbon cycle understanding.

We demonstrate how GeoFoundation models pre-trained on abundant unlabelled Sentinel-3 satellite data can dramatically improve performance when only small amounts of in-situ measurements are available. Our approach first pretrains a model using 512,000 Sentinel-3 tiles spanning global ocean regions to learn generalizable features, then fine-tunes on sparse oceanographic measurements.

The benefits for data-limited applications are substantial. Our primary production model achieved meaningful performance using only 103 in-situ measurements—representing just 6% of the pixels in a single satellite image. When training data was reduced to just 19 observations, the foundation model maintained strong performance whilst conventional approaches failed. This demonstrates the approach's potential to extract maximum value from existing sparse datasets and opportunistically collected measurements.

Beyond improved statistical performance, the model captures realistic spatial patterns over large oceanic regions where no training data exists. Large-scale inference reveals detailed coastal productivity structures that conventional physical models typically under-predict, suggesting the approach has learnt meaningful oceanographic relationships.

We also evaluated the approach for chlorophyll-a concentration estimation using 274 global in-situ measurements. The GeoFoundation model substantially outperformed existing methods, achieving lower RMSE compared to decision tree approaches and the operational Sentinel-3 OLCI Level-2 neural network product. Crucially, when applied to large-scale inference over the North Sea, the foundation model produced spatial patterns with higher Structural Similarity Index Measure (SSIM) (0.88) to the operational product compared to models trained from scratch (0.82), and the decision tree (0.68), demonstrating improved ability to capture realistic oceanographic features.

The implications for operational ocean carbon monitoring are significant. This methodology could enhance existing observation networks by maximising the value of each expensive ship-based measurement and support carbon cycle research in data-poor regions.

An important prerequisite for understanding the role and response of ocean ecosystems to rising atmospheric CO₂ levels and global warming is accurate and well-characterized regional, basin and global scale measurements of oceanic Net Primary Productivity (NPP). Currently global estimates of NPP from satellite data used in global ocean carbon cycling and climate studies, continue to suffer from uncertainties because of their dependence on: 1) satellite derived fields of phytoplankton biomass that are constructed using algorithms incapable of providing product accuracies over regional and global scales, 2) limited estimates of phytoplankton photosynthetic quantum yields (ϕ) that are currently obtained primarily aboard research vessels, and 3) inadequate methods for scaling local in-situ ϕ measurements to regional and basin scales. Here we have utilized the Absorption based Model (AbPM), that exploits the inherent optical absorption properties of phytoplankton derived from remotely sensed reflectance data rather than phytoplankton biomass as an input, and a novel bio-optical classification scheme called Bio-Optical Measurement and Evaluation System (BIOMES) for scaling limited in-situ estimates of ϕ to obtain global maps of NPP. We have used the global collection of in-situ NPP datasets to assess the performance of AbPM derived estimates of NPP against those obtained using more widely used biomass-based models.

The particle size distribution (PSD) is a key property related to the structure and function of marine ecosystems, as well as biogeochemical and optical properties. A recently developed PSD ocean color algorithm retrieves the parameters of an assumed power-law PSD (slope ranging from 2.5 to 6.0, and scaling parameter, N_o) using a 2-component particle bio-optical model, applied to mostly phytoplankton-sized particles and smaller. The PSD can be used to retrieve size-partitioned phytoplankton carbon. Here, we extrapolate the PSD to larger particles to retrieve global mesozooplankton carbon biomass and diel vertical migration (DVM) as part of a larger project aimed at assessing zooplankton contribution to vertical carbon transport via biogenic hydrodynamic transport. A significant finding is that retrieval results are realistic when the PSD slope is fixed at 4.0, rather than variable spatially and temporally within the 2.5 to 6.0 range as in the original PSD retrievals. This fixed slope of 4.0 is consistent with the so-called Sheldon hypothesis; the significance of this finding is discussed, also drawing on recent developments indicating a very conserved marine ecosystem size spectrum. Furthermore, retrievals are realistic when the tuned scaling parameter (N_o) is used, but the tuning was previously designed for POC and phytoplankton carbon retrievals, and it is independent of the mesozooplankton retrievals. The novel algorithm has been applied to the latest monthly OC-CCI v6.0 merged ocean color data set; these data and code have been published. Basic characteristics of this data set are presented and discussed. The PSD-based mesozooplankton carbon biomass and DVM retrievals are validated against two data sets independent of ocean color PSD: a lidar-based DVM retrieval, and the MAREDAT in-situ zooplankton biomass data set. Results of that validation are presented and implications of the time lag between phytoplankton and zooplankton biomass are discussed.