TL;DR
This study analyzes how updates to the Surface Ocean CO2 Atlas (SOCAT) database, which compiles global ocean CO2 measurements, affect estimates of air-sea CO2 fluxes. By using a neural network model trained on five recent SOCAT versions, the researchers found that while newer SOCAT releases generally improve data coverage and reduce uncertainty, a recent decline in data collection since 2017 introduces significant divergence in flux estimates, particularly in key regions like the Southern Ocean.
What Problem Does This Paper Address?
Understanding the ocean's role in the global carbon cycle is crucial for climate modeling and predicting future climate change. A key component of this is quantifying the exchange of carbon dioxide (CO2) between the atmosphere and the ocean, known as air-sea CO2 fluxes. The Surface Ocean CO2 Atlas (SOCAT) is the primary global database for this, compiling in-situ measurements of CO2 in surface seawater. However, SOCAT is continuously updated with new data and undergoes annual quality reassessments. These updates, especially changes in data coverage and the timing of data availability, can significantly influence the calculated global CO2 fluxes. This paper addresses the question of how these annual SOCAT updates, particularly the evolution of data distribution and the impact of recent data collection declines, affect the accuracy and stability of global air-sea CO2 flux estimates.
Methods
The researchers employed a data-driven approach using an ensemble of feed-forward neural networks (FFNNs), specifically the CMEMS-LSCE-FFNN model. This model was trained on SOCAT in-situ CO2 fugacity (fCO2) data and various environmental predictor variables (e.g., sea surface temperature, salinity, wind speed) derived from satellite and reanalysis data. The FFNNs were trained and validated using a 3-month window of data to map global monthly fCO2 fields. For each of the five SOCAT versions analyzed (v2020 to v2024), an ensemble of 50 FFNN models was run. The study then compared the resulting fCO2 estimates, their uncertainties, and the derived air-sea CO2 fluxes (fgCO2) across these different SOCAT versions. Model accuracy was assessed using Root-Mean-Square Deviation (RMSD) against SOCAT data.
Datasets
The study utilized the following datasets:
- SOCAT Data: Five annual releases of the Surface Ocean CO2 Atlas (SOCATv2020, SOCATv2021, SOCATv2022, SOCATv2023, and SOCATv2024) for the period 2000-2023. These datasets represent the in-situ measurements of CO2 fugacity (fCO2).
- Environmental Predictor Variables:
- Sea Surface Temperature (SST), Surface Salinity (SSS), Sea Surface Height (SSH), and Chlorophyll-a (CHL) from the Copernicus Marine Service (CMEMS).
- Mixed Layer Depth (MLD) from the Estimating the Circulation and Climate of the Ocean (ECCO2) Phase II.
- Atmospheric CO2 mole fraction (xCO2) from Copernicus Atmosphere Monitoring Service (CAMS).
- Sea surface fCO2 climatologies (fCOolim) from the Lamont Doherty Earth Observatory (LDEO).
- Wind speed and total dry-air pressure from ERA5.
- Sea ice fraction from CMEMS.
Key Insights and Ideas
- SOCAT Updates Matter: Even seemingly incremental updates to SOCAT, like adding new data or re-assessing existing data, can lead to noticeable shifts in global CO2 flux estimates. This highlights the sensitivity of carbon budget calculations to the underlying observational data.
- The Post-2017 Data Decline is Significant: A noticeable drop in data collection since 2017 has created a divergence in flux estimates between newer and older SOCAT versions, particularly impacting recent years. This suggests that the ocean carbon sink estimates might be less constrained in the most recent period.
- Regional Hotspots are Key: The Southern Ocean, upwelling zones, and the subpolar Pacific are identified as regions where SOCAT data updates have the most pronounced impact on flux estimates. This is likely due to their inherent variability and potentially sparser data coverage.
- Machine Learning is Powerful but Data-Dependent: The FFNN approach effectively maps global CO2 fluxes, but its accuracy is intrinsically tied to the quality and coverage of the input SOCAT data. The study demonstrates how limitations in data can lead to underestimation of CO2 uptake and overestimation of air-sea fluxes.
Key Results
- SOCATv2024 Data Volume: The latest version, SOCATv2024, includes approximately 20% more data than SOCATv2020.
- Global Net CO2 Sink Estimates: SOCATv2024 leads to a revised global ocean net CO2 sink estimate of 1.77 ± 0.17 PgC yr⁻¹, which is 4.5% to 9.6% lower than earlier versions.
- Uncertainty Reduction: Global uncertainty (1σ) improved by 0.3% to 7.6% with SOCATv2024 compared to SOCATv2023 and SOCATv2020.
- Impact of Data Decline: Annual net sink estimates from SOCATv2024 are 0.2 to 1 PgC yr⁻¹ lower than previous estimates, attributed to the decline in data collection since 2017.
- Model Accuracy: RMSD for fCO2 estimates decreased modestly from 20.1 µatm (v2020) to 19.66 µatm (v2024) over the 2000-2023 period, representing a 2.2% improvement.
- Regional Differences: Significant deviations in flux estimates were observed in the eastern North Pacific, parts of the Southern Ocean, the Labrador Sea, and the South Atlantic, with model-data mismatches of about 5-10 µatm.
Limitations and Future Work
- Data Scarcity: The paper acknowledges the persistent challenge of data sparsity in remote regions and specific climate zones, leading to regional biases and difficulties in extrapolating CO2 dynamics accurately.
- Data Latency: A significant delay (at least six months) between data collection and its availability for users is noted, impacting the ability to capture real-time changes.
- Non-Stationary Nature: Temporal and regional CO2 flux variations are subject to non-stationarity, making it challenging to capture emerging trends or short-term fluctuations with limited historical data.
- Future Work: The authors advocate for global efforts to sustain and expand current observation networks, accelerating data collection and dissemination. They suggest prioritizing monitoring key hot-spot regions like the Southern Ocean and subpolar Pacific, and leveraging new technologies (uncrewed surface vehicles, profiling floats, satellite sensors) to enhance data volume and spatial-temporal coverage, especially in remote or harsh environments. Future efforts in data collection, calibration, and mapping methods should proceed with caution to prevent unforeseen system errors and reduce uncertainty.
Summary updated