An improved long-term high-resolution surface pCO₂ data product for the Indian Ocean using machine learning

Area of Science:

• Oceanography
• Climate Science
• Data Science

Background:

• Accurate surface ocean partial pressure of carbon dioxide (pCO2) estimation is vital for understanding the ocean's role in the global carbon cycle and its response to climate change.
• High-resolution model simulations often exhibit deviations from observational data, necessitating correction for reliable climate studies.

Purpose of the Study:

• To enhance the accuracy of surface ocean pCO2 model simulations using machine learning.
• To develop an improved surface pCO2 data product for the period 1980-2019.

Main Methods:

• Employed a machine learning algorithm (XGBoost) to correct deviations between model simulations (pCO2model) and observations (pCO2obs).
• Trained the model to generate spatio-temporal deviations and added them back to the original model simulations.
• Compared the improved data product with moored observations and other gridded datasets (SOCAT, CMEMS-LSCE-FFNN, OceanSODA).

Main Results:

• The machine learning approach improved surface pCO2 data product accuracy by approximately 40% ± 3.31% in RMSE compared to existing datasets.
• Adding climatological deviations yielded greater improvements than adding interannual deviations.
• Demonstrated significant enhancement in model-simulated surface pCO2 outputs.

Conclusions:

• Machine learning algorithms are effective in correcting and improving high-resolution surface ocean pCO2 model simulations.
• The developed data product offers a more accurate representation of surface ocean pCO2, crucial for climate change research.
• Climatological corrections are particularly effective in enhancing model accuracy for surface ocean pCO2.