Surface ocean CO2 concentration and air-sea flux estimate by machine learning with modelled variable trends

Abstract

The global ocean is a major sink of anthropogenic carbon dioxide (CO2) emitted into the atmosphere. Machine learning has been actively used in the past decades to estimate the oceanic sink, but it is still a challenge to obtain an accurate estimate due to scarcely available CO2 measurements. One of the methods to deal with data scarcity was normalizing multiple years’ CO2 values to a reference year to increase the spatial coverage. The practice assumed a constant CO2 trend for the normalization. Here, we used three machine learning models to extract variable ocean CO2 trends on a decadal scale and proposed a method to use the extracted ocean CO2 trends to correct the decadal atmospheric CO2 trends for data normalization. The method minimizes assumptions of using the extracted ocean CO2 trends directly. Comparisons of our CO2 flux estimate with machine learning products included in Global Carbon Budget 2021 indicates that using the variable trends improved the bias resulted from using a constant trend and that the trends are a critical factor for machine learning methods. Our dataset includes monthly distributions of surface ocean CO2 concentration and air-sea flux in 1980-2020 with a spatial resolution of 1×1 degree.

Magnitude, Trends, and Variability of the Global Ocean Carbon Sink From 1985 to 2018

“…Various approaches have been devised to fill the gaps in this database to create near‐global and temporally complete maps of seawater pCO 2 for nearly the entire open ocean and, for RECCAP2, at monthly resolution from roughly the mid‐1980s to 2018, although some recent approaches have extended these estimates back to before 1960 (Bennington et al., 2022; Rödenbeck et al., 2022). These gap‐filling (or interpolation) techniques include statistical (Rödenbeck et al., 2013), multi‐linear regression (Iida et al., 2021), and various machine learning algorithms (Chau et al., 2022; Denvil‐Sommer et al., 2019; Gloege et al., 2022; Gregor et al., 2019; Landschützer et al., 2014; Zeng et al., 2022). The interpolation step in these models is significant because on average only 1%–2% of the 1° × 1° grid cells at any given month are occupied by actual seawater pCO 2 observations, and the remaining 98%–99% must be filled in by the algorithms (Fay et al., 2021; Rödenbeck et al., 2015).…”

Magnitude, trends, and variability of the global ocean carbon sink from 1985-2018

“…resolution from roughly the mid-1980s to 2018, although some recent approaches have extended these estimates back to before 1960 (Bennington et al, 2022;Rödenbeck et al, 2022). These gap-filling (or interpolation) techniques include statistical (Rödenbeck et al, 2013), multi-linear regression (Iida et al, 2021), and various machine learning algorithms (Chau et al, 2022;Denvil-Sommer et al, 2019;Gloege et al, 2022;Gregor et al, 2019;Landschützer et al, 2014;Zeng et al, 2022). The interpolation step in these models is significant because on average only 1%-2% of the 1° × 1° grid cells at any given month are occupied by actual seawater pCO 2 observations, and the remaining 98%-99% must be filled in by the algorithms (Fay et al, 2021;Rödenbeck et al, 2015).…”

An Assessment of CO2 Storage and Sea‐Air Fluxes for the Atlantic Ocean and Mediterranean Sea Between 1985 and 2018

“…These products are based on the interpolation of in situ pCO 2 data accessed from different releases (v2019‐v2021, v5) of SOCAT (Bakker et al., 2016) to near‐global coverage. Several interpolation methods are used including machine learning techniques (Chau et al., 2022; Gloege et al., 2021; Gregor & Gruber, 2021; Gregor et al., 2019; Iida et al., 2021; Landschützer et al., 2014; Watson et al., 2020; Zeng et al., 2022) and a diagnostic mixed layer scheme (Rödenbeck et al., 2013). Sea‐air CO 2 fluxes (FCO 2 ) are computed from reconstructed pCO 2 fields following: $$${\text{FCO}}_{2}=\text{Kw}\,\left(1-{\mathrm{f}}_{\text{ice}}\right)\,{\mathrm{K}}_{0}\,\left({\text{pCO}}_{2}\,-\,{\text{pCO}}_{2},\text{air}\right)$$$ where Kw is gas transfer velocity; f ice is sea‐ice cover fraction; K 0 is CO 2 solubility in seawater; and pCO 2 , and pCO 2 , air are the partial pressures of CO 2 in seawater (nominally at 5 m depth) and in the overlying atmosphere, respectively.…”

“…These products are based on the interpolation of in situ pCO 2 data accessed from different releases (v2019-v2021, v5) of SOCAT (Bakker et al, 2016) to near-global coverage. Several interpolation methods are used including machine learning techniques (Chau et al, 2022;Gloege et al, 2021;Gregor & Gruber, 2021;Gregor et al, 2019;Iida et al, 2021;Landschützer et al, 2014;Watson et al, 2020;Zeng et al, 2022) and a diagnostic mixed layer scheme (Rödenbeck et al, 2013). Sea-air CO 2 fluxes (FCO 2 ) are computed from reconstructed pCO 2 fields following:…”