The increasing concentration of atmospheric CO2 since the onset of the industrial era has been affecting the natural climate due to the greenhouse effect. This effect is partially mitigated by the global ocean CO2 uptakes, which account for about one-quarter of the anthropogenic CO2 emissions. Natural climate variability and anthropogenic climate change also feedback to influence the sea-air CO2 exchange. It is essential to quantify the global ocean carbon sink and its temporal variability to understand further the response of the carbon cycle to future global change. The surface ocean partial pressure of CO2 (pCO2) measurements from the SOCAT dataset were widely used and mapped into continuous gridded data to estimate the sea-air CO2 flux. Due to a lower spatial decorrelation length scale of hundreds of kilometers in the surface ocean than that of thousands of kilometers in the atmosphere, surface ocean pCO2 has more notable spatial variability than atmospheric pCO2. Considerable variability and sparse measurements of surface ocean pCO2 indicate insufficient observations to estimate CO2 flux in most ocean areas directly. Great uncertainty in carbon sink estimation arises from sparse and uneven pCO2 measurements, the gas transfer velocity, and the cool skin effect. Recent application of machine learning algorithms in pCO2 mapping methods increased data availability and further reduced the uncertainty in pCO2-based carbon sink estimates. The average net global ocean carbon sink during the last three decades was documented as −1.40 to −2.45 PgC yr−1. The differences between results were caused by differences in algorithms, division of global biogeochemical provinces, and selection of pCO2 predictors. The accuracy of pCO2 mapping based on machine learning methods remains to be improved, especially in polar regions with sparser pCO2 measurements. The Southern Ocean south of 35°S was a strong carbon sink and has contributed to about 40% of global ocean anthropogenic CO2 uptakes from 1870 to 1995. Changes in the Southern Ocean carbon sink strongly affect the global ocean CO2 uptake. However, the Southern Ocean carbon sink estimated by pCO2-based machine learning methods was about 0.4 PgC yr−1 stronger than the result from global ocean biogeochemical models since 2012. A notable seasonal variability of surface ocean pCO2 was reported in the Southern Ocean, mainly south of 50°S, with high pCO2 levels and carbon sources observed in winter. The strongly seasonally uneven surface ocean pCO2 measurements with missing winter observations may result in an overestimation of the Southern Ocean carbon sink from pCO2 products compared to the in situ observations. Besides supplying more measurements from sailboats or floats, whether the

Numerous studies have utilized machine learning methods for surface ocean pCO2 mapping and CO2 flux estimation, primarily relying on non-linear relationships between SOCAT CO2 measurements (or converted pCO2) and environmental variables. However, the uneven distribution of SOCAT measurements across seasons, particularly in the Southern Ocean, has been a concern. Previous research has highlighted the seasonal variability of surface ocean pCO2 in the Southern Ocean, especially south of 50°S, with higher pCO2 levels and carbon sources observed during winter. This seasonal imbalance in measurements has been implicated in overestimating the Southern Ocean carbon sink in pCO2-based studies compared to in situ observations and biogeochemical models. While efforts to improve pCO2 mapping accuracy using machine learning have been made, uncertainties remain, especially in polar regions with sparse data. Discrepancies between pCO2-based estimates and biogeochemical model results for the Southern Ocean carbon sink have been noted since 2012, further emphasizing the need for improved methodologies and data coverage.

This study employed a boosting ensemble learning feed-forward neural network (BEL FFNNs) method to map surface ocean pCO2 and estimate sea-air CO2 flux. The method used the Surface Ocean CO2 Atlas version 2023 (SOCAT v2023) dataset, which includes quality-controlled global observations of in situ surface ocean fCO2, sea surface temperature, and salinity. The gridded fCO2 data was converted to pCO2 using in situ sea surface temperature and atmospheric pressure. The BEL FFNNs consisted of three FFNNs, where the output of the first FFNN was used as a predictor in the subsequent FFNNs. This architecture aimed to improve the accuracy of pCO2 prediction. The pCO2 predictors were selected using an updated stepwise FFNN algorithm, which iteratively adds or removes predictors based on changes in prediction error. The study compared the performance of the BEL FFNNs and individual FFNNs under different training strategies: using SOCAT pCO2 measurements from all months or only from sectional winter measurements (April-October). The predicted pCO2, root mean square error (RMSE), and bias were compared to evaluate the influence of uneven seasonal measurements. The final pCO2 product integrated pCO2 data from October to April (using all measurements) and May to September (using sectional winter measurements and corrected predictors). The sea-air CO2 flux was estimated using the pCO2 difference across the air-sea interface, accounting for the gas transfer velocity, cool skin effect, and uncertainties in temperature, salinity gradients, and pCO2 reconstruction.

The study's key findings revealed a significant underestimation of surface seawater pCO2 in the Southern Ocean south of 50°S during winter due to seasonally uneven SOCAT measurements. The lack of sufficient winter measurements in the training data led to poor prediction of high pCO2 values during winter months. This resulted in an overestimation of the Southern Ocean carbon sink. Using BEL FFNNs with corrected pCO2 predictors and training with measurements only from April to October substantially mitigated this underestimation. The corrected average Southern Ocean carbon sink during 1992-2021 was -0.87 ± 0.16 PgC yr−1, a decrease of approximately 16% compared to the original estimate (-1.01 PgC yr−1). The overestimation was more pronounced after 2010, with a 29% overestimation in the decadal average carbon sink. The underestimation of pCO2 was most notable south of 60°S, an area often covered by sea ice. The corrected BEL FFNNs product showed better consistency with observations from the Southern Ocean Flux station (SOFS) and the Southern Ocean Carbon and Climate Observations and Modeling (SOCCOM) dataset. Considering the high surface ocean pCO2 under sea ice in winter, complete sea ice melt could further reduce the Southern Ocean carbon sink by approximately 0.14 PgC yr−1. After the Southern Ocean correction, the global ocean carbon sink was estimated to be lower than other pCO2 products but more consistent with the average results from the Global Carbon Budget study. The contribution of the Southern Ocean south of 35°S to global ocean CO2 uptake decreased from -63% in 1992 to 45% in 2021.

This research addresses the significant discrepancy between Southern Ocean carbon sink estimates derived from pCO2 products and biogeochemical models. By identifying and correcting for the underestimation of winter surface ocean pCO2 caused by uneven measurements, the study provides a more accurate assessment of the Southern Ocean's role in global carbon cycling. The findings highlight the importance of adequate seasonal data representation in machine learning models for accurate carbon sink estimations. The corrected estimates are more consistent with results from biogeochemical models, reducing the discrepancy previously observed. The significant impact of sea ice melt on future carbon sink capacity is also underscored, suggesting that climate change effects on sea ice extent will further alter the Southern Ocean's carbon sequestration role. These findings have crucial implications for refining global carbon budgets and predicting the ocean's future response to climate change.

This study demonstrated that the uneven distribution of SOCAT pCO2 measurements between summer and winter led to a substantial overestimation of the Southern Ocean carbon sink over the past three decades. The developed winter correction method, employing a boosting ensemble learning feed-forward neural network with revised predictor selection and training periods, significantly improved the accuracy of pCO2 mapping and carbon sink estimation. The findings highlight the importance of considering seasonal variability and improving data coverage in future studies to accurately assess the Southern Ocean's contribution to global carbon uptake. Future research should focus on improving data collection, particularly during winter months in under-sampled regions, to further refine carbon sink estimates and better understand the impact of climate change on the Southern Ocean's carbon cycle.

The study's primary limitation is the reliance on the SOCAT dataset, which has inherent limitations in spatial and temporal coverage, particularly concerning winter measurements in the Southern Ocean. The correction method focuses on mitigating bias introduced by uneven seasonal sampling; however, other sources of uncertainty in CO2 flux estimations, such as gas transfer velocity and the cool skin effect, remain. The assumption of complete sea ice melt for estimating the impact on the carbon sink is a simplification. The indirect effects of sea ice melt, such as changes in sea surface temperature and convective overturning rates, were not explicitly modeled and could influence the accuracy of the estimate. The spatial resolution of the analysis may also impact the accuracy of the regional estimates, particularly in areas with high spatial heterogeneity.