Trends in Sea-Air CO2 fluxes and sensitivities to atmospheric forcing using an extremely randomized trees machine learning approach

Abstract

Monthly global sea-air CO2 flux maps are created on a 1° by 1° grid from surface water fugacity of CO2 (fCO2w) observations using an extremely randomized trees (ET) machine learning technique (AOML-ET) over the period 1998–2020. Global patterns and magnitudes of fCO2w from AOML-ET are consistent with other machine learning methods and with the updated climatology of Takahashi et al. (2009, https://doi.org/10.1016/j.dsr2.2008.12.009). However, the magnitude and trends of sea-air CO2 fluxes are sensitive to the treatment of atmospheric forcing. In the default configuration of AOML-ET, the average global sea-air CO2 flux is −1.70 PgC yr−1 with a negative trend of −0.89 ± 0.19 PgC yr−1 decade−1. The large negative trend is driven by a small uptake at the beginning of the record. This leads to increasing sea-air fCO2 gradients over time, particularly at high latitudes. However, changing the target variable in AOML-ET from fCO2w to sea-air CO2 fugacity difference, ∆fCO2, results in a lower negative trend of −0.51 PgC yr−1 decade−1, though the average flux remains similar at −1.65 PgC yr−1. This trend is close to the consensus trend of ocean uptake from machine learning and models in the Global Carbon Budget of −0.46 ± 0.11 PgC yr−1 decade−1 switching to a gas transfer parameterization with weaker wind speed dependence reduces uptake by 60% but does not affect the trend. Substituting a spatially resolved marine air CO2 mole fraction product for the zonally invariant marine boundary layer CO2 product yields greater influx by up to 20% in the industrialized continental outflow regions