the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 06 Oct 2025
Reconstruction and Spatiotemporal Analysis of Global Surface Ocean pCO₂ Considering Sea Area Characteristics
Abstract. The partial pressure of carbon dioxide (pCO2) on the surface of the ocean is crucial for quantifying and evaluating the ocean carbon budget. Insufficient consideration of the effects at the sea area scale makes it difficult to comprehensively evaluate the spatiotemporal distribution characteristics and variation patterns of pCO2. This study constructed a pCO2 evaluation dataset based on LDEO measurement data and multi-source data. After conducting correlation testing on a global, far sea, and near sea scale, a ocean surface pCO2 evaluation model was constructed using multiple linear regression, convolutional neural network, gated recurrent unit, long short-term memory network, generalized additive model, extreme gradient boosting, least squares boosting, and random forest. Performance evaluation indicates that the random-forest model consistently achieves the best accuracy across all spatial scales, yielding a global RMSE of 6.123 μatm and an R² of 0.986. In the open ocean, RMSE decreases to 4.699 μatm and R² rises to 0.988, whereas in coastal waters RMSE increases to 8.044 μatm and R² declines to 0.972. Based on this, the annual sea surface pCO2 distribution of 0.25° × 0.25° from 2000 to 2019 was reconstructed. The reconstructed field shows a typical equatorial high/polar low pattern, as well as an overall upward trend consistent with independent observations, with acceleration particularly evident in specific regions of subtropical coastal oceans.
- Preprint
(5684 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2025-4792', Anonymous Referee #1, 03 Nov 2025
The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2025/egusphere-2025-4792/egusphere-2025-4792-RC1-supplement.pdfCitation: https://doi.org/
10.5194/egusphere-2025-4792-RC1 -
AC1: 'Reply on RC1', Yunlong Ji, 08 Nov 2025
Point by point response
Major Points：
Thank you for your very constructive and detail comments concerning our manuscript.
Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval.
Introduction, Line55: It is necessary to clarify the scientific issues existing in previous studies on pCO2.
Response: The clarification content has been marked in red font in the article ‘Through summarizing previous research, it has been found that existing achievements mostly focus on independent analysis of local sea areas or zones, lacking a global perspective, and modeling methods do not fully consider the interactions between sea areas, which affects the accuracy of overall assessment.’
Line 94: pCO2600 µatm is considered an outlier in the article, but sufficientliterature support or physical mechanism explanation is not provided.
Response: Corrected, please see lines 98, Corresponding references have been added.
- Results and discussion: During the discussion, it is suggested to supplement some references and compare and discuss the results of this paper with those of previous studies.
Response: Corrected, please see lines 300,we have undertaken a thorough comparison of our model outputs with the datasets from Zhong et al. (2022) and the Copernicus Marine Serviceproduct.
- Line 150: The transition between Section 3.1 (Correlation Detection) and Section 3.2 (Model construction and evaluation) feels somewhat abrupt. To enhance the logical flow, it would be helpful to briefly state at the beginning of Section 3.2 how the findings from the correlation analysis informed the subsequent modeling step. Response: Done, please see lines 195.
- Line 275: Describe missing and blank values in multi-source data
Response: Done, please see lines 283-286.
- Line 290: To better showcase the novelty of your work, please add a direct comparison with the cited studies (Zhong et al., 2022; Chau et al., 2021)..
Response: Corrected, please see figure 9.
Line 320: Regarding the description of Figure 12, do the influencing factors of PCO2in nearshore areas take into account river inputs or anthropogenic CO2emissions?
Response: We fully agree that river input and anthropogenic CO2 emissions are key processes affecting the carbon cycle in nearshore waters. In the global scale modeling framework of this study, due to the significant regional heterogeneity of the above process and the lack of continuous and consistent observational data support on a global scale, it was not included as an independent driving factor in the random forest model. It should be noted that the biogeochemical parameters (such as pH, chlorophyll concentration, etc.) used in this model as comprehensive environmental indicators have indirectly responded to environmental disturbances caused by river inputs and human activities. Therefore, the reconstruction results of the model in nearshore areas have to a considerable extent reflected the comprehensive effects of these local processes.
- Line 376:The text beginning at line 376 should be moved to a new "Conclusion" section. As this content serves as the concluding discussion for the entire study.
Response: Done, please see lines 389.
Some minor suggestions：
- Line 14: It is recommended to correct the indefinite article for grammatical accuracy. "a ocean surface..." should be changed to "an ocean surface..."
Response: Done, please see lines 14.
- Figure1: It is suggested to supplement the longitude and latitude
Response:Thank you for your suggestion. However, during the revision process, we have attempted to overlay latitude and longitude grids with scale markings. However, the research area has a large span and dense sub regions, and the newly added values significantly obscure the details of the original data and route information. As a result, the map tends to be cluttered and the readability significantly decreases.
- Figure3: The variable represented by the horizontal coordinate needs to be marked
Response:Thank you for the reviewer's suggestion. We have examined Figure 3, where the x-axis represents the Spearman correlation coefficient (ρ). This statistic is a dimensionless indicator defined within the [-1,1] interval, used to measure the strength and direction of monotonic relationships between variables. Therefore, according to the prevailing display standards in this field, physical units are usually not labeled.
- Line 120: “d represents the level difference of the variable”, d should be corrected to D.
Response: Done, please see lines 126.
- Figure 3: o2 in the coordinate axis needs to be corrected
Response:Done, please see figure 3.
-
AC1: 'Reply on RC1', Yunlong Ji, 08 Nov 2025
-
CC1: 'Comment on egusphere-2025-4792', Yanfang Liu, 22 Dec 2025
This article is based on an innovative multi model machine learning framework, which constructs a multi-scale analysis system covering global, offshore, and nearshore areas, and successfully reconstructs surface ocean pCO ₂ data from 2000 to 2019. The research design concept is clear, the technical route is rigorous, and the multi-scale analysis system constructed breaks through the limitations of traditional single scale research. It provides important data support and methodological references for the assessment of ocean carbon cycle and carbon sink potential, and has high academic value and application prospects.
Overall, this article is a solid and meaningful research work, but there are still several details that need to be further optimized and improved before the manuscript is officially published. Specific suggestions are as follows:
1. Line 14: Please revise the indefinite article to "an" to make the phrase "an ocean surface …" grammatically correct.2.Line 119:Section 2.4, Equation (3): The expression "d represents the level difference of the variable" is ambiguous. In addition, there is an inconsistency between the symbol in the text description and that in the equation (text uses d, while the equation uses D) (corresponding to lines 118–120 of the document).
3.Line 185:Figure 3: Please correct the axis label "o2" to "O2".
4.Table 5 (Coastal Model Performance): The RMSE of RF in coastal waters is 8.044 μatm. Although it outperforms other models, this value is nearly double that in the open sea (4.699 μatm). Could the impacts of riverine input or anthropogenic CO₂ emissions be considered?
5.Section 2.3.2: Measured data with pCO₂ > 600 μatm are identified as outliers. However, the full citation of the referenced "previous research experience (19)" is not provided, nor is there any literature or physical mechanism to support this threshold.
6.Section 3.1.2, Table 2: An extra punctuation mark "､" exists at the end of the impact factor list under "Near sea", resulting in non-standard formatting (corresponding to line 178 of the document).
7.Line 275: Please supplement the description of the distribution ratio, processing principles of missing values and blank values in the multi-source dataset, as well as the potential impacts of these values on the results.
Citation: https://doi.org/10.5194/egusphere-2025-4792-CC1 -
AC2: 'Reply on CC1', Yunlong Ji, 23 Dec 2025
Point by point response
Major Points：
Thank you for your very constructive and detail comments concerning our manuscript.
Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval.
1.Line 14: Please revise the indefinite article to "an" to make the phrase "an ocean surface …" grammatically correct.
Response: Done.
2.Line 119:Section 2.4, Equation (3): The expression "d represents the level difference of the variable" is ambiguous. In addition, there is an inconsistency between the symbol in the text description and that in the equation (text uses d, while the equation uses D) (corresponding to lines 118–120 of the document).
Response: Done.
3.Line 185:Figure 3: Please correct the axis label "o2" to "O2".
Response: Done.
4.Table 5 (Coastal Model Performance): The RMSE of RF in coastal waters is 8.044 μatm. Although it outperforms other models, this value is nearly double that in the open sea (4.699 μatm). Could the impacts of riverine input or anthropogenic CO₂ emissions be considered?
Response: We fully agree that river input and anthropogenic CO2 emissions are key processes affecting the carbon cycle in nearshore waters. In the global scale modeling framework of this study, due to the significant regional heterogeneity of the above process and the lack of continuous and consistent observational data support on a global scale, it was not included as an independent driving factor in the random forest model. It should be noted that the biogeochemical parameters (such as pH, chlorophyll concentration, etc.) used in this model as comprehensive environmental indicators have indirectly responded to environmental disturbances caused by river inputs and human activities. Therefore, the reconstruction results of the model in nearshore areas have to a considerable extent reflected the comprehensive effects of these local processes.
5.Section 2.3.2: Measured data with pCO₂ > 600 μatm are identified as outliers. However, the full citation of the referenced "previous research experience (19)" is not provided, nor is there any literature or physical mechanism to support this threshold.
Response: We have corrected the text and added the corresponding references.
6.Section 3.1.2, Table 2: An extra punctuation mark "､" exists at the end of the impact factor list under "Near sea", resulting in non-standard formatting (corresponding to line 178 of the document).
Response: We thank you for your careful scrutiny of Table 2. After re-examining every character, we did not find an extra punctuation mark at the end of the “Near sea” factor list. As a precaution, we have nonetheless double-checked and standardized the entire table to align perfectly with the journal’s formatting guidelines.
7.Line 275: Please supplement the description of the distribution ratio, processing principles of missing values and blank values in the multi-source dataset, as well as the potential impacts of these values on the results.
Response: The blank values are mainly due to the systematic exclusion of land pixels and the limitations of data acquisition in high latitude sea areas: the former is excluded because it does not participate in ocean processes, while the latter is due to the lack of satellite data for key parameters caused by sea ice coverage or insufficient light, resulting in the inability to reconstruct the values in the region.
Citation: https://doi.org/10.5194/egusphere-2025-4792-AC2
-
AC2: 'Reply on CC1', Yunlong Ji, 23 Dec 2025
-
RC2: 'Comment on egusphere-2025-4792', Anonymous Referee #2, 06 Jan 2026
This manuscript presents a multi-scale reconstruction of global surface ocean pCO₂ (2000–2019) using LDEO observations and multi-source environmental data. Multiple statistical and machine-learning models are compared, with random forest (RF) identified as the best-performing method. The resulting 0.25° × 0.25° pCO₂ product reproduces known spatial patterns and temporal trends.
The topic is relevant and the dataset is extensive; however, several issues related to citation accuracy, clarity of scientific contribution, methodological description, and consistency in analysis need to be addressed before the manuscript can be considered for publication.
Major comments
1. The Introduction contains citation mismatches (e.g., line 35: Telszewski et al. cited as Qiu et al., 2022) that should be corrected. More importantly, while many previous studies are listed, the manuscript does not clearly identify the key limitations of existing pCO₂ reconstructions nor explain explicitly how the present multi-scale approach addresses these issues.
2. In Section 2.3.1, the proportion and structure of missing data in the original datasets are not reported. It is unclear whether missing values are sparse or occur in long consecutive gaps, which directly affects the reliability of nearest-neighbour interpolation.
3. The study considers 25 potential predictors, several of which are strongly correlated or physically redundant (e.g., to vs. thetao, ar vs. ca, chl vs. kd490). Multicollinearity of variables might affect the model interpretability. Is there any criterion for retaining or excluding variables?
4. At line 153, “p-value” appears to be used where the Spearman correlation coefficient (ρ) is intended.
5. In Figure 4, the x-axis label “Sample size” is unclear, as no sampling or subsampling experiment is described in the text. In addition, the legend format in Figure 7 should be made consistent with Figures 5 and 6 to facilitate comparison.
6. Language should be double-checked. For example, in line 221, the term “the model” is used without specifying which model is being discussed.
7. Sections 3.1–3.3 already provide detailed model performance metrics. In Section 3.4, additional accuracy statistics (e.g., line 284) are reported without clearly explaining how they differ from earlier results (e.g., independent validation versus internal testing).Please Clearly distinguish internal model evaluation from independent validation of reconstructed products and avoid redundant reporting.
8. The descriptions of machine-learning models (CNN, LSTM, GRU, RF, XGBoost, LSBoost) are largely conceptual. Critical implementation details—such as network architectures, hyperparameters, feature normalization, and optimization procedures—are missing. How are the training/validation/testing splitted?
Citation: https://doi.org/10.5194/egusphere-2025-4792-RC2 -
AC3: 'Reply on RC2', Yunlong Ji, 08 Jan 2026
Point by point response
Major Points：
Thank you for your very constructive and detail comments concerning our manuscript.
Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval.
- The Introduction contains citation mismatches (e.g., line 35: Telszewski et al. cited as Qiu et al., 2022) that should be corrected. More importantly, while many previous studies are listed, the manuscript does not clearly identify the key limitations of existing pCO₂ reconstructions nor explain explicitly how the present multi-scale approach addresses these issues.
Response: Done. Revision of Citations
We have corrected the citation of Telszewski et al. in Line 35. To ensure the completeness and accuracy of all citations throughout the manuscript, we have thoroughly checked the entire reference section, verifying that each citation aligns with the corresponding literature and that the formatting is consistent with academic standards.
(1) Insufficient Consideration of Spatial Heterogeneity
Most existing studies either focus on a single local sea area (e.g., the North Atlantic, Gulf of Mexico, Baltic Sea) or adopt a unified global modeling framework, neglecting the significant differences in environmental conditions, driving factors, and pCO₂ variation characteristics between far sea areas and near sea areas.
To address this issue, our study constructs a multi-scale analysis framework covering the global ocean, far sea areas (water depth > 200 meters), and near sea areas (water depth ≤ 200 meters). The research areas are divided into far sea areas and near sea areas based on water depth, and scale-specific pCO₂ evaluation models are established. For the environmentally stable far sea areas, we emphasize capturing long-term temporal dependencies and signals of large-scale hydrological and biological processes. For near sea areas affected by various complex factors, we incorporate region-specific driving factors and optimize the model structure to adapt to high variability. This targeted approach effectively improves the fitting accuracy and adaptability of the models in different sea area types.
(2) Inadequate Adaptability Between Models and Driving Factors
Existing studies mostly adopt fixed model structures or globally unified combinations of driving factors, failing to fully consider the requirements of environmental complexity differences in different sea areas for model adaptability. Additionally, the selection of driving factors lacks targeting, making it difficult for the models to accurately capture the core impact mechanisms of pCO₂ in different regions.
We resolve this limitation through the comprehensive optimization of models and driving factors: we compared eight machine learning models and identified the Random Forest (RF) model as the optimal model across all scales. Its advantage in capturing complex nonlinear relationships enables it to adapt to the environmental characteristics of different sea areas. Meanwhile, based on Spearman correlation analysis and the SHAP (SHapley Additive exPlanations) method, we screened key driving factors for each scale (e.g., Total alkalinity in sea water (talk) serves as the secondary key factor at the global scale, while the contribution rate of mole concentration of dissolved molecular oxygen in sea water (O₂) significantly increases in near sea areas), ensuring the rationality and targeting of driving factor selection.
(3) Low Reconstruction Resolution
Some existing studies lack the overall processing of spatiotemporal differences in multi-source data, resulting in low spatial resolution of pCO₂ reconstruction products (mostly 1°×1° or coarser), which makes it difficult to accurately reflect the spatiotemporal variation characteristics of pCO₂ within small scales.
We address this limitation through high-resolution and high-precision reconstruction strategies: by processing multi-source data (including strict data matching, outlier handling, and data balancing strategies), we reconstructed the annual pCO₂ distribution with a high resolution of 0.25°×0.25° from 2000 to 2019. The results demonstrate that the accuracy of pCO₂ reconstruction is significantly improved compared with existing studies.
- In Section 2.3.1, the proportion and structure of missing data in the original datasets are not reported. It is unclear whether missing values are sparse or occur in long consecutive gaps, which directly affects the reliability of nearest-neighbour interpolation.
Response: In terms of data types, in-situ observation data (e.g., ar, ca, pH, talk) have a relatively high missing-value percentage (12.29%—12.90%), satellite observation data (chl, kd490) have a missing-value percentage of 7.8%—7.85%, numerical model data (e.g., so, thetao, uo, vo, mlotst*) have a low missing-value percentage (1.41%—4.32%), and satellite observation data related to wind fields (uwind, vwind) have a missing-value percentage of only 0.03%. The overall data integrity is good. All missing values in the dataset stem from limitations in data collection in high-latitude sea areas, mainly due to sea ice coverage and insufficient light restricting satellite remote sensing observations and on-site sampling, resulting in the lack of key parameter data required for pCO₂ reconstruction.
- The study considers 25 potential predictors, several of which are strongly correlated or physically redundant (e.g., to vs. thetao, ar vs. ca, chl vs. kd490). Multicollinearity of variables might affect the model interpretability. Is there any criterion for retaining or excluding variables?
Response: The criteria for retaining variables have been supplemented.
For variables with strong mutual correlations identified in the interaction detection (Section 3.1.1) (e.g., to vs. thetao, ar vs. ca, chl vs. kd490), we did not apply arbitrary exclusion. Instead, their retention is justified by their distinct physical significances: to (sea water temperature) reflects the in-situ surface temperature of seawater, while thetao (sea water potential temperature) accounts for pressure effects—retaining both captures temperature dynamics across ocean layers, which is crucial for simulating CO₂ solubility under varying hydrostatic pressure conditions. ar (aragonite saturation state in sea water) and ca (calcite saturation state in sea water) arise from seawater carbonate equilibrium but respond differently to changes in pH and total alkalinity, enhancing the model’s ability to resolve subtle chemical shifts that regulate pCO₂. chl (mass concentration of chlorophyll a in sea water) directly indicates biological activity (e.g., phytoplankton photosynthesis), while kd490 (volume attenuation coefficient of downwelling radiative flux in sea water) reflects optical properties (e.g., turbidity, light penetration)—together, they provide independent constraints on the biological and physical processes governing pCO₂.
- At line 153, “p-value” appears to be used where the Spearman correlation coefficient (ρ) is intended.
Response: Done.
- In Figure 4, the x-axis label “Sample size” is unclear, as no sampling or subsampling experiment is described in the text. In addition, the legend format in Figure 7 should be made consistent with Figures 5 and 6 to facilitate comparison.
Response: Done.
To clarify the data presentation logic of Figure 4, we have supplemented an explanatory note in its title: Given the large volume of data, plotting all data points would result in visual clutter. Therefore, we randomly selected representative data points to illustrate the performance of different models. The selected data fully cover the global ocean, far sea, and near sea scales, as well as low, medium, and high pCO₂ ranges.
We appreciate your attention to the consistency of figure formats. Figure 7 (Independent verification performance of the models in the near sea areas) has been revised, and its legend structure—including model abbreviations, full names, and the right-axis label "Normalized probability density of model residuals"—is fully consistent with Figure 5 (Global Ocean) and Figure 6 (Far sea areas), ensuring the consistency of comparison logic.
The slight differences in the color bar (color scheme of scatter points) are a deliberate design aimed at better distinguishing the three spatial scales (global, far sea, near sea) while maintaining the same color mapping principle (kernel density is represented by color depth, with darker colors indicating higher concentration of data points). This design does not alter the information structure of the legend or the physical meaning of the data.
- Language should be double-checked. For example, in line 221, the term “the model” is used without specifying which model is being discussed.
Response: Done.
We have modified it to "the constructed surface pCO₂ models" to clearly indicate that it refers to all eight comparative models (including MLR, CNN, GRU, etc.) built for the near sea areas in Section 3.2.3.
- Sections 3.1–3.3 already provide detailed model performance metrics. In Section 3.4, additional accuracy statistics (e.g., line 284) are reported without clearly explaining how they differ from earlier results (e.g., independent validation versus internal testing).Please Clearly distinguish internal model evaluation from independent validation of reconstructed products and avoid redundant reporting.
Response: Done.
The internal model evaluation and the independent validation of reconstructed products are not redundant, as they serve distinct roles:
(1) Internal Model Evaluation (Sections 3.2.1–3.2.3)The core objective of this section is to select the optimal model: by comparing the fitting performance of eight machine learning models (MLR, CNN, GRU, LSTM, GAM, XGBoost, LSBoost, RF) across different spatial scales (global ocean, far sea, near sea), the optimal model for each scale is identified (the RF model ultimately performs best across all scales).The data basis is the processed dataset described in Section 2.3.3, which is randomly split into a training set, validation set, and testing set at an 8:1:1 ratio. All three subsets are derived from the same integrated dataset (LDEO in-situ measurements + multi-source influencing factors).The core metrics focus on "model-to-data" fitting accuracy, including MAE, MAPE, MSE, RMSE, and R² of each model on the training, validation, and testing sets (Table 3–5). For example, the RF model at the global scale achieves a testing set RMSE of 6.123 μatm and R² of 0.986. These metrics only reflect the model’s ability to fit and generalize to similar structured data, without involving product validation in real marine environments.
(2) Independent Validation of Reconstructed Products (Section 3.4)The core objective of this section is to verify product reliability: targeting the 0.25°×0.25° resolution (2000–2019) sea surface pCO₂ reconstructed products generated by the RF model, this section evaluates their applicability and prediction accuracy in real marine environments.The validation data adopts external datasets completely independent of the internal evaluation dataset, including unused LDEO in-situ measurements (not involved in model training/testing) and publicly available observation data such as the Hawaii Ocean Time Series (HOT). These datasets have inherently different characteristics from the internal evaluation dataset.The core metrics focus on "product-to-reality" simulation accuracy, also reporting MAE, MAPE, MSE, RMSE, and R². However, these metrics are derived from the comparison between reconstructed products and independent validation data (e.g., global-scale independent validation RMSE = 19.901 μatm, R² = 0.816), which reflects the application value of the products in complex real environments and is not redundant with the internal evaluation metrics.
- The descriptions of machine-learning models (CNN, LSTM, GRU, RF, XGBoost, LSBoost) are largely conceptual. Critical implementation details—such as network architectures, hyperparameters, feature normalization, and optimization procedures—are missing. How are the training/validation/testing splitted?
Response: We would like to clarify that although hyperparameter tuning of algorithms is not the core focus of this study, it is crucial for the model results. We have supplemented relevant explanations as follows:
Regarding the hyperparameter tuning of machine learning algorithms, we adopted standard tuning strategies and parameter ranges widely accepted in the field of oceanographic parameter estimation. The purpose of this tuning is to ensure the basic stability and reliability of each model, rather than conducting innovative exploration or comparative analysis of tuning methods. Through control experiments, we verified that within a reasonable range of hyperparameters, the relative performance ranking of the eight models remains consistent, and the optimal status of the Random Forest (RF) model across all scales is not affected by minor parameter adjustments. This confirms that the research conclusions (e.g., the superiority of the RF model, the spatiotemporal variation characteristics of pCO₂) do not depend on specific hyperparameter combinations, further supporting that hyperparameter tuning is not the focus of this study.
-
AC3: 'Reply on RC2', Yunlong Ji, 08 Jan 2026
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 273 | 123 | 35 | 431 | 24 | 25 |
- HTML: 273
- PDF: 123
- XML: 35
- Total: 431
- BibTeX: 24
- EndNote: 25
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1