the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 14 Aug 2025
UEx-L-Eddies: Decadal and global long-lived mesoscale eddy trajectories with coincident air-sea CO2 fluxes and biogeochemical conditions
Abstract. Mesoscale eddies are prevalent features within the global ocean that modify the physical, chemical and biological properties as they move and evolve. These modifications can alter the air-sea exchange of CO2, and therefore these features may be hotspots for enhanced or reduced CO2 uptake compared to the surrounding environment. The understanding of the global and regional effect of mesoscale eddies on ocean CO2 uptake is however limited and largely based on single eddies or small regional subsets. Here, we provide a global dataset of 5996 long lived eddies trajectories (lifetimes greater than a year) with corresponding air-sea CO2 fluxes all tracked using a Lagrangian approach between 1993 to 2022. The trajectories comprise 3244 anticyclonic (‘warm core’) and 2752 cyclonic (‘cold core’) eddies and the dataset provides the biogeochemical conditions, including the CO2 fluxes, within and outside each eddy. The dataset refines a previous regional methodology with a focus on climate quality environmental parameters and uses a global neural network for estimating the fugacity of CO2 in seawater (fCO2 (sw)) along with a comprehensive air-sea CO2 flux uncertainty budget. These refinements provide a robust foundation for studying the modulation of air-sea CO2 fluxes by mesoscale eddies. As an example use of the dataset, we investigate the role of mesoscale eddies in modifying the global and regional air-sea CO2 fluxes, by comparing the eddy driven air-sea CO2 flux to that of the surrounding environment. We find that globally, long-lived anticyclonic eddies enhanced the CO2 sink by 4.5 ± 2.8 % (95 % confidence), while long-lived cyclonic eddies reduce the CO2 sink by 0.7 ± 2.6 %. Collectively, the long-lived eddies indicate an enhancement of the ocean CO2 sink by 2.7 ± 1.1 Tg C yr-1. Propagating the air-sea CO2 flux uncertainties was found to be a key component needed to fully understand apparent differences between previous regional and global studies. The long lived eddies (UEx-L-Eddies) dataset is available on Zenodo at https://doi.org/10.5281/ZENODO.16355763 (Ford et al., 2025).
- Preprint
(2125 KB) - Metadata XML
-
Supplement
(307 KB) - BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on essd-2025-463', Yiming Guo, 07 Sep 2025
-
RC2: 'Comment on essd-2025-463', Anonymous Referee #2, 04 Oct 2025
This manuscript presents a global spatiotemporal database of ocean CO2 covering the period 1993–2022, comprising 5,996 long-lived mesoscale eddies (3,244 anticyclonic and 2,752 cyclonic, each with a lifetime >1 year). Surface ocean fCO2 (sw) was estimated using an improved neural-network approach (UExP-FNN-U), and air–sea CO2 fluxes were calculated with FluxEngine, including a comprehensive uncertainty assessment. The study addresses a critical gap in air–sea carbon flux research by focusing on eddy-scale processes, combining satellite-derived eddy trajectories, reanalysis products, and machine-learning estimates. However, the current version of the manuscript spends too much effort interpreting the results and their implications, while paying insufficient attention to the methods and the demonstration of data reliability. Key details regarding data processing, quality control, and the neural-network training procedure are not clearly described. Moreover, the chlorophyll-based experiments and their comparison or validation are inadequate. I strongly recommend that the authors substantially strengthen the validation and methodological sections. Without a more rigorous demonstration of data and method reliability in capturing eddy features, the manuscript is not yet suitable for publication in Earth System Science Data.
Major comments:
- The dataset is mainly based on an existing neural-network model, with no clear modification or adaptation for eddy-specific environments. The authors should clarify why this model is suitable for mesoscale eddy applications and provide targeted validation to demonstrate its reliability. How does this approach differ from other machine-learning models, such as Landschützer’s framework? What specific features make it appropriate for eddy conditions? Without clear evidence that the model captures eddy-related processes or outperforms general models, its applicability to eddy environments remains unconvincing. Is there any better performance between this data product and other data products like Landschutzer, Gregor, Chau data product?
- The comparison in Section 4.2 with previous studies (e.g., Guo & Timmermans 2024; Li et al. 2025; Keppler et al. 2024; Ford et al. 2023) is too general and does not sufficiently explain the sources of divergence in reported results. To strengthen the discussion, I recommend adding a comparative table or supplementary figure that systematically summarizes key methodological differences across studies—including eddy identification criteria, lifetime and radius thresholds, fCO2 estimation methods, temporal coverage, spatial domain, and uncertainty treatment—and clarifying how these factors may drive discrepancies in both magnitude and interpretation. This would provide readers with a clearer comparative framework.
- The authors should devote greater effort to demonstrating the reliability of the data rather than focusing excessively on result interpretation. The current evaluation of the fCO₂ neural network relies only on overall bias and RMSD, which may obscure regional or seasonal biases. A more thorough validation—such as stratified tests by region, season, eddy lifetime, eddy size, or chlorophyll level, and spatial maps of residuals—would better reveal systematic errors and the contribution of different eddy features. Such analyses are essential to improve uncertainty characterization and strengthen confidence in the derived air–sea CO₂ flux estimates.
- Since ESSD focuses on data production and transparency, it would greatly help readers if the authors clearly present the framework of the neural-network architecture as well as the workflow of data processing, testing, and validation. A schematic of machine learning method and a flowchart of data production would substantially improve clarity and reproducibility.
- The analysis focuses only on long-lived eddies (lifetime > 1 year; radius > 30 km), which account for merely ~0.4% of all eddies in the AVISO data. The authors should justify this restrictive sampling choice. Because the vast majority of shorter-lived, smaller-scale eddies—particularly prevalent in western boundary currents and equatorial regions—are excluded, yet they may exert a substantial and possibly different cumulative influence on air–sea CO2 fluxes. To avoid overgeneralization, I recommend that the authors (i) clearly state that their conclusions apply only to this subset of long-lived eddies, (ii) provide the eddy lifetime distribution and grouped statistics (e.g., sample size, mean radius, spatial coverage, flux contribution) across different lifetime classes, and (iii) if feasible, perform a simplified analysis on short-lived eddies, or otherwise explicitly discuss the likely direction and magnitude of biases introduced by their exclusion with reference to existing literature.
- Some data-processing procedures are insufficiently described, which is not sufficient for a data-oriented paper. The authors note that on some days environmental data (e.g., CCI-SST) are missing due to the absence of defined eddy polygons, but the extent of this missing data has not been quantified, nor is it clear whether certain eddies were excluded or flagged. It is recommended to describe the treatment strategy for cases with substantial data loss (e.g., exclusion, interpolation, or flagging). If no exclusion was applied, the potential direction of bias should be discussed. This would ensure that database users can properly interpret and filter eddy time series that may be affected by data quality issues.
- The authors provide a second fCO2 estimate that includes chl-a as an input, but this product is only available from 1997 onward and contains gaps in polar regions during winter. The manuscript currently mentions these limitations only briefly, without quantifying their impact. I recommend that the authors discuss how this affects the temporal and regional representativeness of the results; and (ii) present spatial difference maps or regional statistics comparing the estimates with and without chl-a, to allow readers to assess the role of biological factors in different oceanic regimes.
- The discussion on data reliability and methodological limitations is insufficient. The authors should explicitly identify the main sources of bias arising from both their approach and data, and propose strategies to improve robustness and reproducibility. For example, conducting additional machine-learning experiments to trace and quantify potential biases would help strengthen the credibility of the dataset.
Minor comments:
- I suggest adding a table summarizing the data sources and key characteristics of all variables used in this study (e.g., variable, units, period, resolution, product name, and references). Such a table would improve clarity, allow readers to quickly assess the datasets employed.
- Regarding xCO2 (MBL), please clarify how the meridional band product was mapped onto the 1° field (e.g., through band replication, interpolation, or another approach). Providing this detail would improve the transparency of the data processing procedure.
- The dataset spans 1993–2022. It is recommended that the authors at least comment on decadal-scale variations. For example, they could assess whether the impact of eddies on air–sea fluxes remained stable during the 1990s, 2000s, and 2010s, or if notable changes occurred. Even if a full trend analysis cannot be conducted in the main text, it would be helpful to provide simple decade-wise statistics (e.g., median changes per decade) in the supplementary materials to give readers an initial view of long-term evolution.
- The manuscript states that SOCAT data have been gridded (monthly 1°) and used for model training and testing, but the strategy for splitting the training and test sets is not clearly described. It is recommended that the authors provide details on the partitioning method employed, such as leave-one-time-out or leave-one-location-out cross-validation, or any other approach used, to clarify the reliability and independence of the model evaluation.
- Figure 1 currently only shows the eddy formation points and cumulative flux scatter. It is recommended to also overlay the eddy occurrence frequency or sample density at each grid point (or provide this in a supplementary figure). This would help assess whether certain large flux values are driven by a few exceptionally large or long-lived eddies.
- Although the manuscript notes that anticyclonic eddies enhance CO2 uptake while cyclonic eddies reduce it, it is recommended that the authors further quantify the asymmetry between the two eddy types. For instance, seasonal or regional statistics of flux differences could be provided, along with a discussion of potential physical drivers (e.g., temperature, stratification) and biological drivers (e.g., productivity).
- This study frequently refers to the “weighted mean,” but the weighting scheme is not clearly specified (e.g., latitude, area, or distance weighting). It is recommended that the authors add a statement clarifying how the weights are calculated.
- Figure 1 shows that many eddy formation points are located near the coast or continental shelf, but the manuscript does not specify how eddy polygons overlapping with land on a given day are handled. This could affect the extraction of sea surface variables and flux estimates. It is recommended that the authors clarify the treatment in the Methods section (e.g., applying a land mask to retain only the ocean portion, or excluding eddy days where the overlap with land exceeds a certain threshold). If no such treatment was applied, it would be helpful to provide in the supplementary materials statistics on the number or fraction of eddies overlapping with land.
- This study employs Type II regression to compare the NN and SOCAT data, but it does not briefly explain why Type II rather than ordinary least squares (OLS) regression was chosen. It is recommended to add 1–2 sentences in the statistical comparison section to justify this choice (e.g., Type II regression is more appropriate when both the independent and dependent variables contain measurement errors) and to cite relevant references.
Citation: https://doi.org/10.5194/essd-2025-463-RC2
Data sets
UEx-L-Eddies: decadal and global long-lived mesoscale eddy trajectories with coincident air-sea CO2 fluxes and biogeochemical conditions (v0-2) D. J. Ford et al. https://doi.org/10.5281/ZENODO.16355763
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 1,188 | 45 | 15 | 1,248 | 23 | 26 | 26 |
- HTML: 1,188
- PDF: 45
- XML: 15
- Total: 1,248
- Supplement: 23
- BibTeX: 26
- EndNote: 26
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This manuscript presents a global dataset of long-lived mesoscale eddies (1993–2022) that includes coincident environmental variables, neural-network–based estimates of surface ocean fugacity of CO₂ (fCO₂(sw)), and derived air–sea CO₂ fluxes with comprehensive uncertainty budgets. The dataset builds on the authors’ earlier regional work by integrating a satellite-derived global eddy atlas, reanalysis products, and a refined neural-network methodology (UExO-FNN-U) for estimating fCO₂(sw). Using this global dataset, the authors investigate how long-lived eddies modulate global air–sea CO₂ fluxes and compare their results with other recent estimates obtained using different methods. The findings suggest that anticyclonic eddies tend to enhance the CO₂ sink, while cyclonic eddies slightly reduce it, although the underlying mechanisms remain unclear. Overall, this dataset represents a valuable contribution to the community by improving our understanding of how coherent mesoscale eddies influence air–sea CO₂ exchange. I recommend publication after the following concerns are addressed: