Bayesian correction of satellite-derived pCO2 reveals the biogeochemical impact of hurricanes on ocean–atmosphere CO2 fluxes

The exchange of CO₂ between the ocean and atmosphere is a key process in the global carbon cycle, highly sensitive to disturbances causes by hurricanes and tropical storms. In this study, we applied a Bayesian Recurrent Neural Network (BNNR) to correct satellite-derived surface pCO ₂ estimates from the PISCES-NEMO model in the eastern tropical Pacific, off the coast of Guerrero, Mexico. The model was calibrated using in situ data collected during Tropical Storm Lidia (2017) and supplemented with synthetic cases representing extreme wind conditions. Under fair-weather conditions, the BNNR achieved low errors (MAE = 0.76 µatm; RMSE = 0.99 µatm). Although errors increased under simulated extreme winds, the model successfully captured abrupt increases in pCO ₂ . During Hurricane John (2024), corrected values exceeded 850 µatm, revealing strong CO ₂ degassing pulses that contributed up to one-third of the seasonal air-sea CO₂ flux, a magnitude largely underestimated by deterministic products. By explicitly incorporating uncertainty through credibility intervals, the BNNR provided robust predictions in highly variable conditions. This methodological framework offers a novel tool for improving the monitoring of marine carbon cycle dynamics and for quantifying the biogeochemical impact of extreme events in coastal regions.