Global air-sea CO2 flux inversion based on multi-source data fusion and machine learning

The global gridded dataset of partial pressure of CO2 (pCO2) in the surface ocean and the associated air-sea CO2 flux are crucial for studying climate change and global carbon cycle. However, the complex nonlinear dynamics of atmospheric and marine systems, along with limited observational data bring significant challenges to the inversion of these data. To address these challenges, a two-stage machine learning algorithm was developed. This algorithm incorporates a replacement method for missing ocean data by introducing ocean model simulations to fill these gaps and a machine learning model of dimensionality reduction-clustering-regression to manage system nonlinearity. By integrating in-situ observations, satellite observations and reanalysis datasets, this study reconstructs the global sea surface pCO2 data at monthly 1° × 1°grid from 1993 to 2020, and then derives the corresponding air-sea CO2 flux through the bulk flux formulation. The results demonstrate that the new inversion method can effectively capture the complex relationship between pCO2 observations and other oceanic characteristics data in the surface ocean, allowing for extrapolation to global ocean regions. Compared to other data-based spatio-temporal interpolation methods, the global gridded dataset obtained in this study shows leading performance in terms of root mean square error (RMSE) and the coefficient of determination (R2). Specifically, the average RMSE of the new dataset is reduced by approximately 42 % and 45 % in the Southern Ocean and Arctic Ocean regions comparing with the optimal results from other inversion datasets. Additionally, the new global pCO2 dataset successfully reconstructs the time series close to the observations in coastal and coral reef regions, indicating that the machine learning algorithm can effectively reproduce the time variation characteristics of complex and highly heterogeneous waters. This study successfully applied a multi-source data fusion approach, offering an alternative solution to address the issue of missing ocean observational data, and providing a new perspective for the inversion research of oceanic carbon flux.