Ocean carbon / A sampling experiment


Explore the research

We cannot measure everywhere. With the same number of observations, how much does where we look change what we know?

01 · Introduction

The ocean absorbs our CO₂.
We measure only a sliver of it.

≈ 1/4of human CO₂ emissions is taken up by the ocean each yearSource: Global Carbon Budget
3.4%of monthly 1° ocean grid cells had a ship measurement in 2014
5,000samples per experiment — only where they go changes
Satellite image of a blue and green phytoplankton bloom in the Celtic Sea
A phytoplankton bloom in the Celtic Sea, seen from orbit. The surface ocean is anything but uniform. NASA Earth Observatory

Machine learning turns sparse measurements into complete maps of ocean CO₂. If the measurements are clustered, do the maps get worse — and where should the next ones go?

02 · Background

Where ships have looked,
and where they haven’t.

Each dot is a 1° cell with real surface-water fCO₂ measurements (SOCAT v2025). Drag the globe and switch year or month to see the gaps.

Drag to rotate · pinch or Ctrl + scroll to zoom

Surface-water fCO₂ · µatm
≤200400≥600
LandNo observation

—

Instruments being lowered from a research ship into the sea
A research ship lowers instruments into the sea. Real measurements are costly, sparse and follow ship tracks. NOAA Ocean Exploration

03 · Method

A simulated ocean
as the answer key.

The real ocean has no complete answer. So we use an Earth-system-model ocean as truth, hide part of it, and test how well each sampling rule recovers it.

  1. TruthIPSL-CM6A-LR gives complete monthly surface pCO₂
  2. HideSet aside 20% of month-cells as one shared test set
  3. SamplePick 5,000 month-cells with one rule
  4. LearnOne gradient-boosting model fills the map from SST, salinity, position and month
  5. ScoreCompare with the hidden truth across 20 paired seeds

Main comparison: south of 60°N · connected ocean · area-weighted · 2005, 2010, 2014 · 20 paired seeds

Three rules, the same 5,000 samples

Illustrative maps and block counts: 2005, seed 0, original full domain before the ocean mask. Not the main comparison above.

Random

Every cell has an equal chance

Historical-density

Weighted by 1990–2004 SOCAT density — not a replay of ship routes

Coverage

Prioritises under-sampled places and months

A second experiment: add to the real network

  1. Keep the real networkEvery 2005 SOCAT month-cell that matches the model ocean stays in training
  2. Add 1,000 cells two waysAt random, or only where SOCAT never measured in 1990–2005 — same months, same count
  3. Score the same hidden cellsOne shared test set, five paired seeds; the added cells are never tested

Open-ocean interior (marginal seas and land-adjacent cells removed) · south of 60°N · 2005 · 5 seeds

See for yourself: where samples go, where errors grow

Drag to rotate · pinch or Ctrl + scroll to zoom

Show
Rule

2005 full-domain supporting view (before the ocean mask), not the headline evaluation.

04 · Findings

Three findings

Finding 1

Sampling by historical density
produces higher error.

+27.3%higher mean absolute error than random
RandomHistorical-density

MAE in µatm, lower is better. Each row compares both rules on the same test set, averaged over 20 paired seeds. Worse in every year and every one of the 15 whole-block units.

Finding 2

Spreading out trims extreme errors,
not typical ones.

−1.0RMSE · µatm+0.4MAE · µatm
MAE (typical error)RMSE (extreme error)

Coverage − random, µatm; lower is better. Most of the RMSE gain comes from marginal seas and coastal cells; in the open ocean it nearly vanishes.

Finding 3

Adding 1,000 observations helps a lot.
Aiming them at gaps helps a little.

−2.9RMSE drop from adding 1,000 observations (µatm)−0.13extra drop when they target historical gaps (µatm)
Existing 2005 SOCAT networkNever observed 1990–2005+1,000

RMSE by region (µatm)

Existing network+ random+ gaps

Random − gap RMSE, per seed (µatm; positive favours gaps)

2005, open ocean south of 60°N. Gap targeting helps inside the gaps but slightly hurts where ships already go. Keeping marginal seas (connected ocean) makes the advantage look larger (an extra 1.32 µatm drop), but about 80% of it comes from Hudson Bay; the earlier global version was inflated by the Arctic and inland lakes. These domain-matched reruns were chosen after seeing the global result.

05 · Robustness

Ask the question differently.
Does it still hold?

3/3

Every year

Historical-density is worse in 2005, 2010 and 2014

15/15

Whole blocks held out

Worse in all 15 year × spatial-fold units: +18.0%

Ocean mask

Direction unchanged after removing lakes, marginal seas and coasts

−4.95 → −0.47

Signed bias

Historical − random bias (µatm, whole-block test) shrinks this much once cells are area-weighted and the domain aligned — too sensitive to be a headline

06 · Conclusion

In these tests, historical-density sampling costs accuracy.
Both addition rules improve reconstruction.

Sampling location matters, but these tests do not identify an optimal deployment. Every result is conditional on one simulated ocean and one learner; benefits for the real network remain unverified.

07 · Limits

What this study
cannot tell you.

08 · Sources & credits

Data, code and people

Data

Related work

Course & team

Independent extension of a Spring 2025 EESC/STAT 4243 project in Galen A. McKinley’s class. Original team: Azam Khan, Bokai He, Sarah Pariser and Zhi Wang.

ReconstructOceanCarbonP3G1

Details

Method details, every table and the audit trail live in the notebook and docs.