Ocean carbon products by LDEO

Over the last few years the McKinley Ocean Carbon Group at LDEO has used various machine learning techniques to interpolate available sparse measurements of surface ocean fCO2 to estimate near global coverage of fCO2 fields. From these, monthly air-sea CO2 fluxes can be estimated, and then integrated to quantify the ocean carbon sink. These efforts build on the legacy of Taro Takahashi who developed the technology to make automated fCO2 measurements, compiled the first database of these measurements, and from them, created the first mapped climatological products.

The McKinley group has developed and released two separate observation-based products which are updated annually each summer following the SOCAT database release: fCO2-Residual and LDEO-HPD. While both methods rely on machine learning methods and the SOCAT database of fCO2 observations, they are unique and independent estimates of the global ocean carbon levels. They each create near global coverage of fCO2 estimates at monthly 1x1 degree resolution, span multiple decades, and are updated each year as new observations become available. More information and download links for data and code are available from this website on the tabs at top.

Schematic showing interpolation method

In addition to these two machine learning products, our group has also created a final update to the LDEO Takahashi Climatology. Here we include the updated climatological mean distribution of fCO2 and corresponding net sea–air CO2 flux estimate developed by the group at LDEO, in honor of the legacy of Taro Takahashi. This climatology represents the mean of ocean conditions over the last four decades and is unique relative to other statistical, mechanistic and machine learning approaches in that it interpolates in time and space using only the available fCO2 data rather than using proxy variables for gap filling.

taylor diagram comparing various pCO2 products

Similarities and differences between the methods- a quick glance

LDEO fCO2-Residual approach for fCO2 reconstruction applies pre-processing to remove the direct effect of temperature, simplifying the target variable for machine learning. This method yields an estimate for the observation period, beginning in 1985.

LDEO Hybrid Physics Data (LDEO-HPD) approach uses hindcast model fCO2 output as a prior, or “first guess” estimate and then uses machine learning to estimate global ocean monthly fCO2. Through this method, it is possible to obtain an estimate of fCO2 for years available in model output, typically beginning in the late 1950s, thus allowing for a longer timescale of reconstruction.

Both reconstruction methods use the eXtreme Gradient Boosting (XGB) algorithm method to learn a non-linear relationship between the target variable and observed predictors.

Technical note: The partial pressure of CO2 (pCO2) and fugacity of CO2 (fCO2) are nearly identical, with the latter accounting for the non-ideality of the gas. Historically, the ocean carbon community has done its work in terms of pCO2. However, fCO2 is more accurate and the community is now converting to this focus. On this website, you will see both fCO2 and pCO2.