Visualizing surface ocean CO2 data

Exploring surface ocean CO2 data

The McKinley group and our collaborators in SOCOM use machine learning to turn sparse ship-based measurements of surface ocean CO2 from SOCAT into complete global maps. These maps are key to quantify the ocean carbon sink in the Global Carbon Budget and how we study its mechanisms.

But how sparse are those measurements, really?

Bokai He, a former student in my Spring 2025 EESC/STAT 4243 class, built this globe as part of an independent extension of his class project. Each dot is a 1° grid cell with real surface-water fCO2 measurements from SOCAT v2025. Drag to rotate, pick a year or month, and tap a dot to read its value. The gaps are the story: in 2005, only about 2% of monthly 1° ocean grid cells had a ship measurement; by 2014, sampling improved, but was still only about 3%.

Drag to rotate · pinch or Ctrl + scroll to zoom

Year

Month

Surface-water fCO₂ · µatm

≤200400≥600
LandNo observation

Loading…

Unfortunately, these sampling gains have reversed in recent years. Sampling coverage in the last 5 years (2020-2024) has fallen to only about 2% (Dong et al. 2024).

The McKinley group is collaborating on a new project that will fill some of the most glaring gaps in observations which are found in the Southern Ocean in austral winter (June-July-August). With the Constraining Ocean Carbon with Optimized Observing (COCO2) project, we are using uncrewed surface vehicles (USVs) from Saildrone to go to the dangerous waters of the wintertime Southern Ocean to get these essential data. We think Schmidt Sciences for their generous support for COCO2!

Bokai’s full data story also explains more about machine learning reconstructions and the application of Earth System Models (ESMs) as testbeds to understand the impacts of different sampling strategies. Using an ESMs as a known “truth,” he asks whether where we sample changes how well machine learning can reconstruct the full field. With the same 5,000 samples, allocating them according to historical ship density produced about 27% higher mean absolute error than sampling at random. Adding observations to the real network helps substantially; aiming them at historical gaps helps a little more. This builds on testbed work from our group, including Gloege et al. (2021) and Heimdal et al. (2024). Code and data are on GitHub.

Congratulations, Bokai, on a beautiful piece of work!

Globe and data processing by Bokai He. We thank the SOCAT data providers, researchers, funding agencies, data managers and quality controllers. SOCAT is an international effort endorsed by IOCCP and SOLAS.

– Galen A. McKinley, November 16, 2026