A new observation-based estimate of the Southern Ocean carbon sink, south of 30°S, has been achieved by using machine learning to combine traditional ship-based measurements from SOCATv2025 with year-round data from autonomous Biogeochemical-Argo (BGC-Argo) floats.
Argo floats are one of three observing platforms being used in TRICUSO. These autonomous, vertically-profiling platforms enable continuous monitoring of the ocean’s physical and biogeochemical properties. The global array, currently consisting of approximately 4,000 floats, forms the international Argo programme, which is part of the Global Ocean Observing System (GOOS) and Global Climate Observing System (GCOS). Argo has proposed a global and multidisciplinary Argo Program — the OneArgo mission — to expand and enhance its capabilities by 2030 (Roemmich et al., 2019).”
When fully implemented, the OneArgo array will contain 4,700 floats, of which 1,000 will be Biogeochemical Argo (BGC-Argo). Euro-Argo ERIC,one of the three Research Infrastructures involved in TRICUSO, coordinates the European contribution to this design, with a current foreseen commitment to deploy and manage 25% of the active array.
Despite the progress of the OneArgo mission, notable Argo data gaps persist in polar regions, particularly in the Southern Ocean, where harsh winter conditions and remote high-latitude areas remain chronically undersampled. TRICUSO’s work directly addresses this by providing recommendations of the optimal float network configuration to reduce uncertainty in estimates of air-sea CO2 fluxes.
BGC-Argo floats can be fitted with a range of sensors; among which are those measuring pH, dissolved oxygen, temperature and salinity. These data can be combined with empirical algorithms to derive carbonate system parameters, including surface ocean pCO2 (Sauzède et al., 2017; Williams et al., 2017; Bittig et al., 2018). Different BGC‑Argo float configurations can be utilised in these calculations, reflecting trade‑offs between lower‑cost BGC-Argo floats equipped with dissolved oxygen sensors and the more expensive pH and dissolved oxygen‑equipped systems that provide higher‑accuracy pCO2 estimates. With the further addition of acoustic sensors (see the web story for milestone 12), air-sea CO2 fluxes can be estimated directly from BGC-Argo floats. Although significantly more cost-effective than ship-based campaigns, deploying and maintaining BGC-Argo floats still carries considerable cost, making strategic deployment essential. A central goal of TRICUSO is therefore to determine where, when, how many, and what type of BGC-Argo floats are needed to optimise the observing network, with a focus on the Southern Ocean, hosting the largest share of global ocean CO2 uptake.
Ships and Floats Together: A New Benchmark for Southern Ocean CO₂ Uptake
TRICUSO researchers have delivered a new observation-based estimate of the Southern Ocean carbon sink south of 30°S, using machine learning to combine traditional ship-based measurements from SOCATv2025 with year-round data from autonomous BGC-Argo floats. Because ships struggle to operate in the Southern Ocean’s harsh conditions (particularly during the long austral winter and at high latitudes near the sea-ice edge) ship-based observations are heavily biased toward summer months and mid-latitude shipping routes, leaving critical gaps in space and time (Figure 2). The work, in review for Global Biogeochemical Cycles, shows that the Southern Ocean absorbed approximately 1.32 ± 0.10 billion tonnes of carbon per year between 2003 and 2024 (i.e. about 12% less than estimates based on ship observations alone) and finds that the “strengthening” of the sink reported in earlier ship-only studies largely disappears once winter and high-latitude float observations are included.
This study, led by Louise Delaigue (Sorbonne Université (LOV)), brings together TRICUSO co-authors Peter Landschützer (VLIZ), Cathy Wimart-Rousseau (NOC), Lisandro Arbilla (LOCEAN-CNRS), Jean-Baptiste Sallée (LOCEAN-CNRS), and Hervé Claustre (Sorbonne Université (LOV)), alongside external collaborators from partner institutions. This work directly supports TRICUSO’s work package three objectives of reducing uncertainty in observation-based estimates of ocean carbon uptake by integrating complementary observing systems.
Why It Matters
The Southern Ocean accounts for roughly 35-45% of the global ocean’s uptake of atmospheric CO2, despite covering only about a fifth of the ocean surface. Quantifying this sink accurately is essential for the international assessments that guide climate policy: the annual Global Carbon Budget (Friedlingstein et al., 2025), which tracks where human CO2 emissions end up each year, and the Intergovernmental Panel on Climate Change (IPCC) assessments, which inform projections of future warming and the carbon budgets compatible with the Paris Agreement, both depend on robust estimates of how much CO2 the Ocean is absorbing.
Yet quantifying this sink has remained one of the most stubborn challenges in carbon-cycle science: ship-based measurements compiled in the Surface Ocean CO2 Atlas (SOCAT) are heavily concentrated in summer and along mid-latitude shipping routes, leaving large gaps in winter and at high latitudes, precisely where the carbon-cycle dynamics are most active. As a result, recent published estimates of mean Southern Ocean uptake have ranged nearly twofold, and the trajectory of the sink over recent decades has remained contested. Reducing this uncertainty directly improves the reliability of global carbon accounting and the climate projections built on it.
What TRICUSO Delivered
The team trained a multilayer perceptron (MLP) neural network on three observing-system configurations (SOCAT-only, BGC-Argo float-only, and an uncertainty-weighted combination of both) to produce monthly maps of surface Ocean CO2 partial pressure across the Southern Ocean at 1° × 1° resolution from 2003 to 2024. Air–sea CO2 fluxes were then computed from each reconstruction and compared against independent constraints from the Global Carbon Budget (Friedlingstein et al., 2025), including atmospheric inversions and global ocean biogeochemical models (Figure 3).
Key outcomes include:
- A revised mean Southern Ocean CO2 sink of −1.32 ± 0.10 Pg C yr⁻¹ south of 30°S over 2003-2024, consistent with atmospheric inversion (−1.39 ± 0.05 Pg C yr⁻¹) and biogeochemical model (−1.35 ± 0.08 Pg C yr⁻¹) estimates.
- A 12% reduction in inferred basin-scale uptake relative to SOCAT-only reconstructions, attributable to better representation of wintertime outgassing south of the Polar Front.
- A near-zero, statistically non-significant long-term trend in the combined reconstruction (−0.003 ± 0.002 Pg C yr⁻²), indicating that the strong sink intensification implied by ship observations alone is largely an artefact of evolving observational coverage.
- A framework for anchoring autonomous float observations to ship-based measurements through uncertainty-weighted training, mitigating known systematic biases in float-derived pCO2 while preserving the seasonal and latitudinal coverage that floats uniquely provide.
The findings reinforce a central message of TRICUSO’s observing-system work: neither ships nor floats alone can deliver a robust, basin-scale picture of Southern Ocean carbon uptake. Together, however, they provide a physically consistent benchmark that can be used to evaluate Earth system models, constrain future climate projections, and detect genuine climate-driven changes in Ocean carbon uptake against the backdrop of an evolving observing network. The work also highlights priorities for the next phase of the observing system, including direct pCO2 measurements on autonomous platforms and sustained cross-calibration between ships, floats, and uncrewed surface vehicles.
The full manuscript, “The Southern Ocean pCO2 sink in an evolving observing system” (Delaigue et al., in review for Global Biogeochemical Cycles), is available at https://essopenarchive.org/doi/full/10.22541/essoar.15002449/v1.
Looking Ahead
Building on the combined ship-float reconstruction presented above, which directly addresses the pCO2 products used in the Global Carbon Budget, a natural next step is to extend this evaluation to the Ocean biogeochemical models that the Global Carbon Budget (Friedlingstein et al., 2025), equally relies upon. Work currently in progress addresses this question through a multi-model subsampling assessment using the Self-Organizing Map – Feed-Forward Neural Network (SOM-FFN) method (Landschützer et al., 2013), in which five global Ocean biogeochemical models are sampled at SOCAT and BGC-Argo floats locations across the Southern Ocean.
Ongoing work aims to evaluate reconstruction skill across multiple models and timescales, quantify the added value of integrating BGC-Argo floats alongside ship-based measurements by disentangling whether the improvement comes from having more observations overall or from the unique winter and high-latitude seasonal coverage that floats provide, and identify systematic biases introduced by float-only or ship-only configurations, particularly with respect to long-term trend detection.
Very preliminary results are encouraging and consistent with the findings from the pCO2 reconstruction: combining both observing systems improves skill across all models tested, while float-only configurations show systematic biases in long-term trends. Once the full analysis is complete, this work will provide a robust framework for optimising the future Southern Ocean observing network and further reducing uncertainties in global carbon budget estimates. This work is further complemented by ongoing TRICUSO Observing System Simulation Experiments (OSSEs) that incorporate the economic cost of float deployments, with the aim of identifying optimal network configurations for the Southern Ocean.
Milestone story written by Louise Delaigue (Sorbonne Université (LOV)), Lisandro Arbilla (LOCEAN-CNRS) and Clara Douglas (Euro-Argo ERIC).
References
Bittig, H. C., Steinhoff, T., Claustre, H., Fiedler, B., Williams, N. L., Sauzède, R., Körtzinger, A., Gattuso, J.-P. (2018). An alternative to static climatologies: Robust estimation of open ocean CO2 variables and nutrient concentrations from T, S, and O2 data using Bayesian neural networks. Frontiers in Marine Science, 5, 328. https://doi.org/https://doi.org/10.3389/fmars.2018.00328
Friedlingstein, P., O’Sullivan, M., Jones, M. W., Andrew, R. M., Hauck, J., Landschützer, P., Le Quéré, C., Li, H., Luijkx, I. T., Olsen, A., Peters, G. P., Peters, W., Pongratz, J., Schwingshackl, C., Sitch, S., Canadell, J. G., Ciais, P., Jackson, R. B., Alin, S. R.,…Zeng, J. (2025). Global Carbon Budget 2024. Earth Syst. Sci. Data, 17(3), 965-1039. https://doi.org/10.5194/essd-17-965-2025
Landschützer, P., Gruber, N., Bakker, D. C. E., Schuster, U., Nakaoka, S., Payne, M. R., Sasse, T. P., Zeng, J. (2013). A neural network-based estimate of the seasonal to inter-annual variability of the Atlantic Ocean carbon sink. Biogeosciences, 10, 7793-7815. https://doi.org/10.5194/bg-10-7793-2013
Sauzède, R., Bittig, H. C., Claustre, H., Pasqueron de Fommervault, O., Gattuso, J.-P., Legendre, L., & Johnson, K. S. (2017). Estimates of water-column nutrient concentrations and carbonate system parameters in the global ocean: A novel approach based on neural networks. Frontiers in Marine Science, 4, 128. https://doi.org/10.3389/fmars.2017.00128
Thierry, V., Claustre, H., Pasqueron de Fommervault, O., Zilberman, N., Johnson, K. S., King, B. A., Wijffels, S. E., Bhaskar, U. T. V. S., Balmaseda, M. A., Belbeoch, M., Bollard, M., Boutin, J., Boyd, P., Cancouët, R., Chai, F., Ciavatta, S., Crane, R., Cravatte, S., Dall’Olmo, G.,…Zuo, H. (2025). Advancing ocean monitoring and knowledge for societal benefit: the urgency to expand Argo to OneArgo by 2030 [Review]. Frontiers in Marine Science, Volume 12 – 2025. https://doi.org/10.3389/fmars.2025.1593904
Williams, N. L., Juranek, L. W., Feely, R. A., Johnson, K. S., Sarmiento, J. L., Talley, L. D., Dickson, A. G., Gray, A. R., Wanninkhof, R., Russell, J. L., Riser, S. C., Takeshita, Y. (2017), Calculating surface ocean pCO2 from biogeochemical Argo floats equipped with pH: An uncertainty analysis, Global Biogeochem. Cycles, 31, 591–604. https://doi.org/10.1002/2016GB005541
