Impact of Predictor Variables on Estimates of Global Sea-Air CO2 Fluxes Using an Extra Trees Machine Learning Approach

Abstract

Monthly global sea-air CO2 flux estimates from 1998-2020 are produced by extrapolation of surface water fugacity of CO2 (fCO2w) observations using an Extra-trees (ET) machine learning technique. This new product (AOML_ET) is one of the eleven observation-based submissions to the second REgional Carbon Cycle Assessment and Processes (RECCAP2) effort. The target variable fCO2w is derived using the predictor variables including date, location, sea surface temperature, mixed layer depth, and chlorophyll-a. A monthly resolved sea-air CO2 flux product on a 1˚ by 1˚ grid is created from this fCO2w product using a bulk flux formulation. Average global sea-air CO2 fluxes from 1998-2020 are -1.7 Pg C yr-1 with a trend of 0.9 Pg C decade-1. The sensitivity to omitting mixed layer depth or chlorophyll-a as predictors is small but changing the target variable from fCO2w to air-water fCO2 difference has a large effect, yielding an average flux of -3.6 Pg C yr-1 and a trend of 0.5 Pg C decade-1. Substituting a spatially resolved marine air CO2 mole fraction product for the commonly used zonally invariant marine boundary layer CO2 product yield greater influx and less outgassing in the Eastern coastal regions of North America and Northern Asia but with no effect on the global fluxes. A comparison of AOML_ET for 2010 with an updated climatology following the methods of Takahashi et al. (2009), that extrapolates the surface CO2 values without predictors, shows overall agreement in global patterns and magnitude.