Near-real-time monitoring of global ocean carbon sink

Piyu Ke, Xiaofan Gui, Wei Cao, Dezhi Wang, Ce Hou, Lixing Wang, Xuanren Song, Yun Li, Biqing Zhu, Jiang Bian, Stephen Sitch, Philippe Ciais, Pierre Friedlingstein, Zhu LiuarXiv 2023 · DOI · arXiv

TL;DR

Researchers have developed Carbon Monitor Ocean (CMO-NRT), a deep learning-based framework that provides near-real-time, monthly gridded estimates of the global ocean carbon sink. By using Convolutional Neural Networks (CNNs) to emulate complex biogeochemical models, the team has reduced the typical 1–2 year data reporting lag to just a few weeks, offering a vital tool for tracking the Earth's carbon budget as it happens.

What Problem Does This Paper Address?

The ocean is a massive carbon sponge, absorbing about a quarter of anthropogenic $CO_2$ emissions. However, our ability to monitor this "sink" in real-time has been surprisingly limited.

Currently, global estimates rely on two main sources: Global Ocean Biogeochemical Models (GOBMs), which are computationally expensive and slow to run, and observation-based products (like SOCAT), which suffer from a significant time lag in data collection and quality control. Consequently, the Global Carbon Budget usually reports with a one-year delay, and many models have a two-year latency. This "data void" makes it difficult for policymakers to assess the immediate impact of climate policies or to detect rapid shifts in ocean carbon uptake caused by events like El Niño.

Methods

The authors introduce a machine learning pipeline designed to "fill in" the recent past by learning the relationship between easily observable environmental variables and complex carbon flux outputs.

  • The Architecture: The team employed a stacking approach of multilayer Convolutional Neural Networks (CNNs) and linear models. CNNs are particularly effective here because they treat the global ocean as a grid (image-like data), allowing the model to capture spatial correlations and regional patterns.
  • Semi-Supervised Learning: To improve stability, they used a semi-supervised framework. They trained the model on "patches" (18x18 grids) and used data augmentation (randomly removing features) to calculate an unsupervised loss. This helps the model generalize better even when some input data might be noisy.
  • Predictors: The model uses nine environmental drivers: sea surface temperature (SST), sea ice fraction, salinity, atmospheric $CO_2$ concentration ($xCO_2$), mixed-layer depth, sea surface height, chlorophyll-a, sea level pressure, and wind speed.
  • Bridging the $xCO_2$ Gap: Since even atmospheric $CO_2$ data has a slight lag, they used a LightGBM (Light Gradient Boosting Machine) model to estimate recent $xCO_2$ levels based on historical trends and marine boundary layer data.

Datasets

The model doesn't just predict one "truth"; it emulates the entire ensemble of the Global Carbon Budget 2022.

  • Target Data: Outputs from 10 different GOBMs (e.g., NEMO-PlankTOM12, MPIOM-HAMOCC6) and 8 observation-based data products (e.g., MPI-SOMFFN, Jena-MLS).
  • Input Data: High-resolution satellite and reanalysis products, including NOAA’s OISST, ECMWF’s ERA5 and ORAS5, and ESA’s GlobColour.
  • Temporal Scope: The resulting CMO-NRT dataset provides monthly $1^\circ \times 1^\circ$ resolution data from January 2022 to July 2023.

Key Insights and Ideas

  • Deep Learning as an Emulator: The core "aha" moment is that deep learning can act as a high-fidelity proxy for computationally heavy biogeochemical models. Once trained, the CNN can produce global maps in seconds that would take a supercomputer days to simulate using traditional differential equations.
  • Stability through Ensembling: By training separate models to emulate 18 different established products (10 models + 8 data products), the researchers provide a range of estimates. This preserves the scientific uncertainty inherent in the field while providing a "near-real-time" average.
  • Spatial Context Matters: By processing data in 18x18 patches rather than point-by-point, the model accounts for the fact that carbon flux in one area of the ocean is physically linked to its neighbors.

Key Results

  • High Accuracy: In validation tests (predicting 2020–2021 data using only earlier data), the models achieved an $R^2$ value above 0.9 for most products.
  • Low Bias: The mean difference between the deep learning predictions and the original complex models was generally less than 3 $\mu atm$ for surface ocean $fCO_2$.
  • Seasonal Fidelity: The CMO-NRT successfully captured the strong seasonal cycles of carbon uptake and release across different latitudes, matching the "gold standard" Global Carbon Budget outputs with high precision.

Limitations and Future Work

  • The "Extremes" Problem: The model performs less accurately in regions with extreme $fCO_2$ values, such as the equatorial Pacific (where deep, $CO_2$-rich water upwells) and the Arctic Ocean.
  • Arctic Data Scarcity: Some observation-based products have significant missing data in the Arctic, which limits the CNN's ability to learn reliable patterns in those regions.
  • Future Directions: The authors suggest that as more in-situ sensors (like BGC-Argo floats) become available, integrating that real-time "ground truth" directly into the machine learning framework could further refine the accuracy of these near-real-time estimates.

Data Access: The CMO-NRT dataset is publicly available on Figshare and via the Carbon Monitor Ocean website.

Summary updated

Abstract

Mitigation of climate change will highly rely on a carbon emission trajectory that achieves carbon neutrality by the 2050s. The ocean plays a critical role in modulating climate change by sequestering CO2 from the atmosphere. Relying on the multidisciplinary cutting-edge methodologies and technologies, the near-real-time monitoring of global ocean carbon sinks from January 2022 to July 2023 aims to provide the world's latest assessment of monthly and gridded global ocean carbon sinks based on machine learning and other data science technologies. The project will help us find a robust route to deal with climate change, which will significantly promote the ocean carbon sinks research and will be of great interest for policy makers, researchers, and the public. This research aims to build up an integrated machine learning framework and methodology for assessing global ocean carbon neutral process; development of near-real-time dataset; development of visualization platform; research papers published in international prestigious journals; an executive report openly accessible to policy makers and the public.
BibTeX
@misc{ke2023near,
  author = {Piyu Ke and Xiaofan Gui and Wei Cao and Dezhi Wang and Ce Hou and Lixing Wang and Xuanren Song and Yun Li and Biqing Zhu and Jiang Bian and Stephen Sitch and Philippe Ciais and Pierre Friedlingstein and Zhu Liu},
  title = {{Near-real-time monitoring of global ocean carbon sink}},
  year = {2023},
  eprint = {2312.01637},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2312.01637},
}
Figure