Reconstruction of Global Ocean Surface pCO2 and Air-Sea CO2 Flux: Based on Multigrained Cascade Forest Model

Abstract

Quantifying the role of air-sea CO₂ exchange is essential for accurately estimating the global carbon balance, which is dependent on the spatial and temporal resolution of ocean surface carbon dioxide partial pressure (${p\text{CO}}_{2\left(\text{sw}\right)}$). When dealing with the global ocean as a vast and complex system, most existing studies tend to partition the global ocean into small-scale regions. To account for interactions of environmental variables across multiple regions, we used machine learning algorithms to holistically reconstruct a 20-year global ${p\text{CO}}_{2\left(\text{sw}\right)}$ map at a high resolution of 4 × 4 km based on products from the Moderate Resolution Imaging Spectroradiometer, reanalysis data, and Surface Ocean CO₂ Atlas. Three machine learning methods were compared, with multigrained cascade forest (gcForest) demonstrating the highest accuracy in global reconstruction (r² of 0.92, root mean square error of 13.46, and mean absolute error of 7.34 μatm). The global ${p\text{CO}}_{2\left(\text{sw}\right)}$ has shown a steady increase at an average annual growth rate of 1.95 ± 0.05 μatm yr⁻¹, controlled mainly by sea surface temperature and chlorophyll concentration. This study covers an ocean area of approximately 335 × ${10}^6$ km², encompassing over 95% of the annual average carbon sink area. During 20 years, the daily CO₂ flux decreased by 0.44 mmol m⁻² d⁻¹, while the proportion of carbon sink area remained constant, indicating ocean's carbon uptake capacity per unit area has been increasing.