The Southern Ocean carbon sink has been overestimated in the past three decades

Atmospheric CO2 has risen since the industrial era, with the global ocean taking up roughly one-quarter of anthropogenic emissions. Accurately quantifying the ocean carbon sink and its variability is essential for understanding climate feedbacks. Surface ocean pCO2 measurements (e.g., SOCAT) are sparse and unevenly distributed, especially in the Southern Ocean and during winter, leading to uncertainty in machine-learning-based pCO2 maps and air–sea CO2 flux estimates. A persistent discrepancy exists between Southern Ocean carbon sink estimates from pCO2 products and global biogeochemical models, with products indicating a stronger sink since about 2012. Given reported high wintertime pCO2 and source behavior south of ~50°S, the study tests the hypothesis that seasonally uneven sampling biases machine learning reconstructions toward summer conditions, causing underestimation of winter pCO2 and overestimation of the Southern Ocean carbon sink.

Prior works have applied machine learning and regression to map pCO2 and estimate fluxes, yielding a global sink of roughly −1.40 to −2.45 PgC yr−1, with differences attributed to algorithms, predictors, and regionalization. The Southern Ocean is a dominant anthropogenic CO2 sink but shows strong seasonality and interannual variability; earlier pCO2-based products report a stronger Southern Ocean sink than global ocean biogeochemistry models since ~2012. Studies highlight sparse winter observations, especially south of 50–60°S, and note wintertime pCO2 can exceed atmospheric values, suggesting potential source behavior obscured by sampling gaps. Traditional regression approaches are particularly sensitive to such uneven sampling. Recent efforts (e.g., ensemble ML, SOM-FFNN, OceanSODA-ETHZ, CMEMS) improved uncertainty but still diverge from model-based estimates, with the discrepancy concentrated in the Southern Ocean.

Data: SOCAT v2023 gridded fCO2 was converted to pCO2 using in situ SST and pressure (ERA5 sea-level pressure), applying standard virial coefficients, gas constant, and temperature. Environmental predictors (remote sensing and reanalysis variables; details in Supplementary Table 1) were used. Modeling: A boosting ensemble learning of three feed-forward neural networks (BEL FFNNs) was constructed. Each FFNN used multiple hidden layers (10 neurons per layer) with outputs from FFNN I used as predictors for FFNN II and III. Outputs were averaged over different initializations. Predictor selection: A Stepwise BEL algorithm objectively selected region-specific predictors by iteratively adding/removing variables to minimize prediction error, accounting for co-correlation. Regions followed a self-organizing map regionalization: latitude bands 35–50°S, 50–60°S, and south of 60°S, and three basin sectors for 35–50°S. Winter correction: To mitigate seasonal sampling imbalance, two adjustments were made for months May–September south of 50°S: (1) training only on sectional winter-period SOCAT data (April–October) to rebalance seasonality, and (2) increasing the weighting of winter errors in the predictor-selection criterion so that winter contributes comparably to summer (modified RMSE: 3× weight for May–Sep). Month was also allowed as a predictor. For 35–50°S no correction was applied due to lack of imbalance effects. The final product uses: October–April pCO2 from all-month training; May–September pCO2 from winter-corrected predictors and April–October-trained networks. Validation and error estimation: K-fold cross-validation (by year, K=4) compared RMSE and bias between predicted pCO2 and SOCAT. Independent validation used the SOFS time-series (142.0°E, 46.8°S) and SOCCOM float data. Traditional regressions (MLR, MNLR) were tested for comparison. CO2 flux calculation: Air–sea CO2 flux F = k(α_subskin pCO2w − α_skin pCO2atm), with atmospheric pCO2 from NOAA MBL xCO2 and ERA5 pressure (humidity-corrected), CO2 solubilities from T and S, and transfer velocity k = Γ(660/Sc)^0.5 U^2 using ERA5 winds scaled by Γ = 0.27 to satisfy 13C constraints. Sea ice cover (ERA5) masked air–sea exchange in ice-covered regions. Uncertainty: Major components included transfer velocity (assumed 10%), cool skin effect (3% temperature, 1.7% salinity after subskin correction), and pCO2 mapping/averaging (grid 5 µatm, mapping 7–25 µatm regionally). Combined pCO2 interpolation uncertainty across provinces ≈ ±0.13 PgC yr−1. Total Southern Ocean flux uncertainty: ±18.4% (≈ ±0.16 PgC yr−1).

- Seasonally uneven SOCAT sampling (winter data ~one-fifth of summer) leads to underestimation of wintertime surface pCO2 south of 50°S in ML products, especially south of 60°S. - Winter correction (training on April–October data plus winter-weighted predictor selection) reduces winter bias and RMSE substantially: • South of 60°S, BEL FFNNs trained on all months showed a winter bias of −5.77 µatm; corrected approach reduced bias to −1.38 µatm, with RMSE improvements (e.g., to ~13.29 µatm with corrected predictors and Apr–Oct training). • In 50–60°S, using corrected predictors and Apr–Oct training reduced winter RMSE to 10.93 µatm and bias to −0.25 µatm. • Traditional regressions suffered much larger winter errors (e.g., MLR south of 60°S: RMSE 34.02 µatm; bias −17.29 µatm). - The underestimation of winter pCO2 leads to overestimation of Southern Ocean CO2 uptake: • Over 1992–2021, the average Southern Ocean sink is corrected from −1.01 to −0.87 ± 0.16 PgC yr−1 (≈16% overestimation previously). • Since 2010, overestimation was ≈29% (decadal mean sink reduced from −1.20 to −0.93 PgC yr−1 after correction). • May–September average Southern Ocean uptake decreases by ~0.34 PgC in the corrected product relative to uncorrected. - Spatially, the largest winter ΔpCO2 corrections occur south of 60°S, but sea ice suppresses winter flux there; the principal impact on overestimated sink arises from 50–60°S where sea ice is limited. - Corrected global ocean sink (using corrected Southern Ocean) is lower than many prior pCO2 products and more consistent with Global Carbon Budget model averages, narrowing the product–model discrepancy rooted in the Southern Ocean. - Scenario analysis indicates that complete sea-ice melt would expose high winter pCO2 waters south of 60°S, reducing the annual Southern Ocean sink by ~0.14 PgC yr−1 on average (winter release ~0.28 PgC yr−1), further weakening the global ocean’s CO2 uptake.

The study directly addresses the long-standing discrepancy between pCO2-based products and biogeochemical models for the Southern Ocean by identifying and correcting a key bias source: seasonal sampling imbalance. By rebalancing training data seasonality and emphasizing winter conditions in predictor selection, the winter pCO2 underestimation south of 50°S is markedly reduced, leading to a weaker and more model-consistent Southern Ocean sink estimate. The findings reconcile much of the 2010s divergence and confirm that the main bias arises in the 50–60°S belt where winter fluxes are not masked by sea ice. The corrected interannual variability reproduces the well-known 1990s weakening and post-2001 reinvigoration of the sink, but with a slower strengthening rate after winter correction. The results imply that previous observationally based products likely overstated Southern Ocean uptake, affecting global budgets and trend attribution. They further highlight the sensitivity of flux reconstructions to data distribution and to region-specific drivers (e.g., vertical mixing in winter versus biological drawdown in summer).

This work demonstrates that uneven seasonal sampling in SOCAT, especially the scarcity of winter observations south of 50°S, biases machine-learning pCO2 reconstructions and overstates the Southern Ocean carbon sink by ~16% over 1992–2021 and ~29% since 2010. A boosting ensemble FFNN with winter-focused training and predictor selection reduces winter pCO2 bias, yielding a corrected average Southern Ocean sink of −0.87 ± 0.16 PgC yr−1 and improved agreement with biogeochemical models in the Global Carbon Budget. The largest pCO2 corrections occur south of 60°S, but sea ice limits flux; the net sink overestimation mainly arises from 50–60°S. If Antarctic sea ice were to melt completely, the Southern Ocean sink would decline further by ~0.14 PgC yr−1 on average due to winter outgassing. Future work should (i) expand wintertime observations (ships of opportunity, floats, moorings) particularly under and near sea ice; (ii) refine region- and season-specific predictors capturing winter mixing and circulation; (iii) integrate under-ice and marginal ice zone processes; and (iv) further constrain transfer velocities and cool skin corrections to reduce flux uncertainty.

- Observational sparsity in winter, particularly south of 60°S and under sea ice, necessitates corrective weighting and indirectly constrains validation; under-ice pCO2 remains poorly sampled. - The winter correction relies on training period adjustments and predictor weighting rather than additional independent winter data, and thus may still miss unresolved processes. - Sea-ice melt scenario assumes complete removal and neglects indirect feedbacks (e.g., SST changes, mixing, biology), so flux changes represent a first-order estimate. - Flux uncertainties remain from gas transfer velocity parameterization (assumed 10%), cool skin corrections (temperature 3%, salinity 1.7%), and pCO2 mapping/averaging (regional mapping errors 7–25 µatm; aggregate ~±0.13 PgC yr−1 in the Southern Ocean), yielding a total ±18.4% (~±0.16 PgC yr−1). - Discrepancies with float-derived pCO2 (computed from pH/alkalinity) in certain waters suggest potential biases in indirect estimates that complicate cross-platform validation.