the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 11 Mar 2022
A new estimate of oceanic CO2 fluxes by machine learning reveals the impact of CO2 trends in different methods
Abstract. Global oceans have absorbed a substantial portion of the anthropogenic carbon dioxide (CO2) emitted into the atmosphere. Data-based machine learning (DML) estimates for the oceanic CO2 sink have become an import part of the Global Carbon Budget in recent years. Although DML models are considered objective as they impose very few subjective conditions in optimizing model parameters, they face the challenge of data scarcity problem when applied to mapping ocean CO2 concentrations, from which air-sea CO2 fluxes can be computed. Data scarcity forces DML models to pool multiple years’ data for model training. When the time span extends to a few decades, the result could be largely affected by how ocean CO2 trends are obtained. This study extracted the trends using a new method and reconstructed monthly surface ocean CO2 concentrations and air-sea fluxes in 1980–2020 with a spatial resolution of 1×1 degree. Comparing with six other products, our results show a smaller oceanic sink and the sink in early and late year of the modelled period could be overestimated if ocean CO2 trends were not well processed by models. We estimated that the oceanic sink has increased from 1.79 PgC yr-1 in 1980s to 2.58 PgC yr-1 in 2010s with a mean acceleration of 0.027 PgC yr-2.
- Preprint
(1641 KB) - Metadata XML
-
Supplement
(460 KB) - BibTeX
- EndNote
Status: closed
-
RC1: 'Comment on essd-2022-71', Anonymous Referee #1, 18 Mar 2022
Zeng et al. used three machine algorithms (neural network, random forest, and gradient boosting) to estimate ocean pCO2 on a 1x1 grid from 1980-2020. They trained each algorithm to learn SOCAT fCO2 observations using full-coverage fields (SST, SSS, MLD, CHL, LAT, LON, YEAR) as inputs to each algorithm. The output from these algorithms were averaged to create the final product and a bulk parameterization was used to estimate flux. Their flux estimates were lower than the 6 products used in the global carbon budget 2021.
-
AC1: 'Reply on RC1', J. Zeng, 20 Mar 2022
Thanks to the reader for the comment.
Strictly speaking, YEAR was not included in training machine learning models directly. Instead, the annual increase rates of CO2 at decatal scales were extracted by an iteration mathod and the rates were used to remove the trend in CO2 measurements. The trend removed (or normalzied) data were used for training the models. The iteration method used machine learning to remove the dependence of CO2 on SST, SSS, MDL, CHL, LAT and LON and used linear regression to removed the dependence of CO2 on YEAR.
In addtion to citing the referenc of the iteration method, we will put the description in the next revision.
Citation: https://doi.org/10.5194/essd-2022-71-AC1 -
RC2: 'Reply on AC1', Anonymous Referee #1, 20 Mar 2022
Thank you for the reply. Please also note the supplement material attached to my original comment.
Citation: https://doi.org/10.5194/essd-2022-71-RC2 -
AC2: 'Reply on RC2', J. Zeng, 08 Apr 2022
The reader has made valuable comments. We have recalculated fluxes of all products used in the comparison, compared fluxes of NIES-ML2 obtained by CCMP and ERA5 wind, and revised the manuscript to include the changes. The attached document includes our point-to-point response to the reader’s comments.
-
AC2: 'Reply on RC2', J. Zeng, 08 Apr 2022
-
RC2: 'Reply on AC1', Anonymous Referee #1, 20 Mar 2022
-
AC1: 'Reply on RC1', J. Zeng, 20 Mar 2022
-
RC3: 'Comment on essd-2022-71', Anonymous Referee #2, 13 Apr 2022
Review of
A new estimate of oceanic CO2 fluxes by machine learning reveals the impact of CO2 trends in different methods
By Zeng et al
Summary
The authors present an updated pCO2 data-based machine learning product which specifically focuses on improving the rate of increase estimated by the model approach. Rather than a constant trend correction to collapse available data to a base-year for analysis, they use a variable rate. Further comparisons with other DML approaches which similarly use a fixed trend reflects that the ocean carbon flux estimate would be overestimated at both ends of the time series with such an approach.
While this finding is interesting and important for the community overall, this paper suffers from broad statements that are not supported with literature citing or sufficient analysis as well as numerous grammatical errors. While the product itself is a valuable update/addition to the field, revisions to this manuscript should be required before acceptance.
Broad statements
The explanation of the creation of the product could be improved to make it clearer. Specifically, the choice of outlining the steps 2.1 and 2.3. Additionally, comparisons in figures are made as compared to a fixed-rate estimate, but I would be interested to see comparisons made to the same method if a simple atmospheric xco2 time-varying trend was used instead. Would the improvement be seen with that choice and the rate extraction step not providing much benefit? There isn’t a clear comparison of that and from Figure 1a it seems like it would do pretty well with atmospheric xCO2.
I feel that there are many, many places in this manuscript where the statements are too broad and unclear or over-ambitious. In the abstract itself, where it states, “the result could be largely affected by how ocean CO2 trends are obtained”, this doesn’t make sense because other DML techniques are not assigning or removing “ocean CO2 trends”. They are using atmospheric xCO2 as one (of many) driver variables and the machine learning approach estimates the pCO2 in areas missing data and from that ocean trends can be calculated. Therefore, this type of statement is not appropriate and especially not in the abstract.
Additionally, the statement in the abstract, “…and the sink in early and late year of the modelled period could be overestimates if CO2 trends were not well processed by models” is incorrect. Yes, this product has a smaller sink as compared to the others, but it does not show a difference in both early and late periods as compared to all other products- just those that use a fixed rate to correct observations to a base year. And what is “not well processed by models”? You must be careful about your terminology. Models are different from DML products as discussed in this and similar literature. So, to say that it is not well processed by models just doesn’t make any sense. The models are not processing the data- the machine learning methods are.
Regarding the comparison to other products, specifically Figure 6, first, the layout seems like a waste of space for these figures and it could be condensed more (also include a legend on each figure subplot for ease of interpretation). Second, the authors leave out a comparison to the JENA product which is one of the most long-available pCO2 products and has been included in the GCB for many years now. And lastly, they don’t account for the fact that the products all are calculating flux using their own choice of wind and atmospheric products, gas exchange parameterization and other choices. These choices could easily make the difference in the increase in flux seen in the last decade or so as compared to the presented NIES -ML3 product. For example, if the wind product chosen by CMEMS-FNN has stronger winds in the 2000s, it could result in a larger flux and therefore that would account for some of the difference seen here. Unless you are standardizing for all of these factors (as done in the pySeaFlux package/product ensemble) then you cannot confidently say the difference are due to their choice of method regarding co2 trend handling.
Another overarching question I had when reading through this method is that the authors need to be sure to adequately handle the fact that global location of the observations are totally variable in time and that definitely impacts the trend. There is no way to account for or correct for the fact that some years data is dominated from the tropical pacific which could have a much different trend than that of the rest of the ocean. This could play a large role in the differences seen as well.
The effort that the authors put to explain their choice of rate and how it impacts the model result is substantial and accounts for about half of the manuscript. But in the end, they get a value that is very close to the (smoothed) atmospheric xco2 increasing rate. The comparisons made in Figure 2 are great at showing a single trend is inappropriate for such a long-time scale of data that is now available, but it doesn’t show why this method of rate assessment is an improved method over just using atmospheric xCO2 trends.
In Section 3.2 where the flux is presented, the authors report an uncertainty value. It is important to note that this uncertainty is solely due to the machine learning choice used, and represents a spread around the three methods shown here. There is no uncertainty included from other aspects of the flux calculation such as the flux parameterization for example.
The flux convention used in this manuscript is opposite of that typically used in the observation-based pCO2 product literature. First and formost though, the convention is never defined and explained that a negative flux value would mean an efflux of carbon out of the ocean. I am confused as to why the authors chose this convention in the first place, especially then used a colorbar that is opposite of expected (negative values in warm colors) so that the maps look like the maps in the literature presenting other products. I would highly suggest flipping the convention around, specifically for the maps and discussion in section 3.2.
In regard to Figure 6 and the comparison to other available DML products, it could be that the differences seen in the later years could be influenced by the nearly logarithmic increase in available data for recent decades as opposed to the 1980s and 1990s. I would be interested to hear the authors comment son this in the discussion in Section 3.3.
Specific comments
There are way too many grammatical and clarifying errors to correct and list here. Some common ones are the use of the word “residue” where it likely should be “residual” and “special” where they meant “spatial”.
Line 149: “The YON was about 20 years in early 1990s” as shown in Figure 1b- I don’t see that from Figure 1b at all. It seems that the line has flattened well before 20 years on the x-axis. What makes the author state this value here?
Line 152: An example of where the authors are over-confident in their statements. “One of the reasons must be that the data points after 2000….” Is that really the ONLY reason that could possibly explain this? I don’t think so. A better phrase would be that one of the reasons “could be”…
Line 194: Figure 3 shows the annual fluxes and Figure 4 shows the spatial distributions.
Line 195: Typo- should be Global Carbon Budget 2021.
Figure 4: Include an explanation of what a, b, and c are in the figure caption. If you are going to state that “the patterns agree well with those in GCB-2021, then show this comparison. What models is this comparing to, and which compare well? The models, spatially, actually have a LARGE range of estimates of flux (and thus pCO2), see Fay & McKinley 2021 in GRL. To broadly say that these maps compare well with the GCB2021 models is untrue and unsupported in this paper.
Line 207: Why are you comparing your 1980-2020 to GCB-2021 for years 2011-2020? You show in Figure 6 that these are years where many of the products start to diverge. Why not do the comparison for the same subset of years?
Line 560: “GSB-2021” should be “GCB-2021”.
Citation: https://doi.org/10.5194/essd-2022-71-RC3 -
AC3: 'Reply on RC3', J. Zeng, 18 Apr 2022
The reader has given valuable comments and kindly pointed out many misuses of terminology, grammar and spelling errors. We have corrected the errors and incorporated most advice in the revised manuscript. The followings are point-to-point response to the comments.
- The reader suggested explaining more about the choice outlined 2.1 and 2.3. We have provided python code for model creation in the supplement. Those interested in the details could go on reading related python documents and references listed in the manuscript. In the revised manual script, we added more explanations for the iteration method of rate extraction.
- Regarding the question of why not just using atmospheric xCO2 trends. It sounds simple but is questionable. First, what xCO2 trends to use? Using the annual increase rate of xCO2 in the same year or in the previous year, or the mean rate of the past two years, three years… 10 years? Second, how to prove that ocean CO2 change follows the same pattern as xCO2? The uptake of air CO2 may not the only major factor that leads to ocean CO2 increase. Changing SST could be an important factor as well, for example. For these reasons, we do not plan to use xCO2 trends directly to model flux for comparison.
- We have rewritten the discussions for comparison. First, we included the JENA product. We didn’t include it initially because its spatial and temporal resolutions are different from others. The results of a comparison could depend on how a monthly product in 1x1 degree grids were derived from JENA’s daily product in 2x2.5 degree grids. Seconds, we recalculated the fluxes of all products under comparison using the same method and adjusted the fluxes to the ones as if they have the same grid coverage. Third, time-series fluxes are put in the same plot for comparison. Fourth, we added a figure to summaries the mean differences.
- The reader questioned the variable of global locations of observations. As the reader pointed out “There is no way to account for or correct for the fact that some years data is dominated from the tropical pacific which could have a much different trend than that of the rest of the ocean.” Our method examined the change of APPARENT GLOBAL TREND with data length and select the trend when the change become small. The implicit assumption is that the effect of unbalanced sampling becomes acceptably week with increasing data length. The method is not perfect, but it offers an alternative. If the locations were quite even geographically, it would not be necessary to use machine learning in the first place, as a simple spatial interpolation would be sufficient.
- In the revised manuscript, we emphasize that the uncertainty value in Section 3.2 is solely due to the machine learning. It is unreality to count in all uncertainty factors in such a data paper, including uncertainties of measurements, griding, and etc.
- In the revised manuscript, we explained that a negative flux value means an efflux of carbon out the ocean and a positive flux into the ocean. Although the reader strongly suggests flipping the convention around, we think the convection makes it easier to see the trend of flux. When we made the flux map, we used the same convention and a similar colorbar palette as those in the figure 6 of GCB-2021.
Citation: https://doi.org/10.5194/essd-2022-71-AC3
-
AC3: 'Reply on RC3', J. Zeng, 18 Apr 2022
-
RC4: 'Comment on essd-2022-71', Anonymous Referee #3, 28 Apr 2022
Review for ‘A new estimate of oceanic CO2 fluxes by machine learning reveals the
impact of CO2 trends in different methods’ by Jiye Zeng et alThe paper aims to investigate issues related to interpolating sparse in situ ocean pCO2 data to allow complete fields to be generated and then global assessments of air-sea gas fluxes and the net oceanic sink to be assessed. The study investigates the impact of assumed or specified trends that exist within some of these interpolation methods towards understanding what impact these assumed trends can have on the resultant oceanic net sink estimate (and in particular the results at the ends of time series). The analysis includes producing a novel reconstruction method which produces an integrated net sink value which is markedly different in magnitude to all of the other datasets analysed.
I would recommend that the paper is significantly revised, including a modified title before then being reviewed again. The reasons for this recommendation are given below.
My main comments are:
- The paper seems to be focussed on a trend analysis, and identifying the impact of assumed trends within different datasets, rather than providing ‘a new estimate of oceanic CO2 fluxes’ and determining the absolute oceanic sink. The methods identify how the rate of ocean absorption has changes and how each of the methods that are also compared (within the inter-comparison) does or does not capture similar trends. As a result the paper title seems a bit misleading as the title should really explain that's its focussed on the analysis of the trends and the rate of ocean absorption of carbon, and not the oceanic sink strength and magnitude (and not on presenting a new estimate of the oceanic sink). The focus in the title and abstract also seems to be a bit out of scope for the journal ESSD as I thought that the journal is focussed on providing datasets for wide use, where the focus here seems to be more focussed on novel science or result-focussed work.
- A detailed methods section seems to be missing from the paper. This lack of information makes it hard to evaluate the approach presented and the results. The limited amount of information on methods (which appear in the section called ‘data’) lack detail and need significantly expanding. In particular the methods used to calculate the air-sea fluxes and the derivation of each of the parameters needed as input to the flux calculation are missing (eg do the NOAAs Marine Boundary Layer Reference provide pCO2atmosphere data, or do they provide XCO2 data that you use to calculate pCO2atmosphere?) The calculation of the net integrated air-sea fluxes is also missing (eg how were the values in Figure 6 calculated, what land and ocean masks were used, how were mixed pixels containing land and ocean handled, was surface area calculated using assuming a sphere or an ellipsoid?). Providing a separate section on methods that detail all of the methods used would help the reader. I could not follow the methods in section 2.3 Rate extraction. A more detailed overview is really needed in the paper for this part of the methods (The authors rely on the reviewer/reader reading another paper to gain a basic understanding).
- The difference between air-sea flux data products is likely to be at least partly (maybe mostly) driven by the differences in input data for calculating the air-sea gas fluxes (eg differences in wind speed data, or sea surface temperature, salinity etc) and this could also be one of the causes for the different identified trends. This issue has been overlooked and should at least be discussed and/or tested for. It may be possible to partially remove it, eg by using a common method to calculate the gas fluxes, a common set of wind data, common atmospheric data and then just the pCO2water fields from each of the datasets within the inter-comparison. This would allow the authors to isolate the impact of different wind and atmospheric forcing data.
- All of the spatially complete datasets used within the inter-comparison form submissions to the Global Carbon Budget assessments. Within the GCB assessments these datasets are all assessed using SOCAT data. If the authors want to keep the ‘new estimate of oceanic CO2 fluxes’ as a key finding and focus of the paper, then their new pCO2 fields should be evaluated using the community SOCAT database eg train on a subset of the dataset, test against the remainder (eg RMSD and bias), or at the very least evaluate their complete pCO2 dataset using the data that were used to train the approach. At the moment the paper presents an inter-comparison where the new approach is vastly different from all other data included in the inter-comparison but no justification or evaluation as to why these new data are useful is given; only that that they are different and new. Can the authors provide some sort of statistical assessment or evaluation of the output data (pCO2 fields and/or gas fluxes) to provide the reader with some confidence that the new results are valid?
General comments
The paper would benefit from the help of a copy editor help improve the main text.
Citation: https://doi.org/10.5194/essd-2022-71-RC4 -
AC4: 'Reply on RC4', J. Zeng, 02 May 2022
The reviewer has given valuable opinions, including changing the title to focus on the dataset. Yes, our intention is to improve our previous method for ocean CO2 reconstruction, not the focus on scientific discovery. We are considering to change the title to “A surface oceanic CO2 product reconstructed by using machine learning to extract CO2 trends at decadal scale”. The following are point-to-point to the reviewer’s comments.
- We are adding more details in the method section to address the issues of air-sea flux calculation, rate extraction, and etc.
- Instead of using flux values in comparison products, we used their pCO2 to recalculate fluxes by a common method. Discussions on the comparison have been revised substantially.
- Our method did set aside part of SOCAT fCO2 for validation. We called it “leave-one-year-out” validation method. For each of the three machine learning methods, 41 validations were done by set aside data in 1980 to 2020 year by year for validation and use the rest for training.
Citation: https://doi.org/10.5194/essd-2022-71-AC4
Status: closed
-
RC1: 'Comment on essd-2022-71', Anonymous Referee #1, 18 Mar 2022
Zeng et al. used three machine algorithms (neural network, random forest, and gradient boosting) to estimate ocean pCO2 on a 1x1 grid from 1980-2020. They trained each algorithm to learn SOCAT fCO2 observations using full-coverage fields (SST, SSS, MLD, CHL, LAT, LON, YEAR) as inputs to each algorithm. The output from these algorithms were averaged to create the final product and a bulk parameterization was used to estimate flux. Their flux estimates were lower than the 6 products used in the global carbon budget 2021.
-
AC1: 'Reply on RC1', J. Zeng, 20 Mar 2022
Thanks to the reader for the comment.
Strictly speaking, YEAR was not included in training machine learning models directly. Instead, the annual increase rates of CO2 at decatal scales were extracted by an iteration mathod and the rates were used to remove the trend in CO2 measurements. The trend removed (or normalzied) data were used for training the models. The iteration method used machine learning to remove the dependence of CO2 on SST, SSS, MDL, CHL, LAT and LON and used linear regression to removed the dependence of CO2 on YEAR.
In addtion to citing the referenc of the iteration method, we will put the description in the next revision.
Citation: https://doi.org/10.5194/essd-2022-71-AC1 -
RC2: 'Reply on AC1', Anonymous Referee #1, 20 Mar 2022
Thank you for the reply. Please also note the supplement material attached to my original comment.
Citation: https://doi.org/10.5194/essd-2022-71-RC2 -
AC2: 'Reply on RC2', J. Zeng, 08 Apr 2022
The reader has made valuable comments. We have recalculated fluxes of all products used in the comparison, compared fluxes of NIES-ML2 obtained by CCMP and ERA5 wind, and revised the manuscript to include the changes. The attached document includes our point-to-point response to the reader’s comments.
-
AC2: 'Reply on RC2', J. Zeng, 08 Apr 2022
-
RC2: 'Reply on AC1', Anonymous Referee #1, 20 Mar 2022
-
AC1: 'Reply on RC1', J. Zeng, 20 Mar 2022
-
RC3: 'Comment on essd-2022-71', Anonymous Referee #2, 13 Apr 2022
Review of
A new estimate of oceanic CO2 fluxes by machine learning reveals the impact of CO2 trends in different methods
By Zeng et al
Summary
The authors present an updated pCO2 data-based machine learning product which specifically focuses on improving the rate of increase estimated by the model approach. Rather than a constant trend correction to collapse available data to a base-year for analysis, they use a variable rate. Further comparisons with other DML approaches which similarly use a fixed trend reflects that the ocean carbon flux estimate would be overestimated at both ends of the time series with such an approach.
While this finding is interesting and important for the community overall, this paper suffers from broad statements that are not supported with literature citing or sufficient analysis as well as numerous grammatical errors. While the product itself is a valuable update/addition to the field, revisions to this manuscript should be required before acceptance.
Broad statements
The explanation of the creation of the product could be improved to make it clearer. Specifically, the choice of outlining the steps 2.1 and 2.3. Additionally, comparisons in figures are made as compared to a fixed-rate estimate, but I would be interested to see comparisons made to the same method if a simple atmospheric xco2 time-varying trend was used instead. Would the improvement be seen with that choice and the rate extraction step not providing much benefit? There isn’t a clear comparison of that and from Figure 1a it seems like it would do pretty well with atmospheric xCO2.
I feel that there are many, many places in this manuscript where the statements are too broad and unclear or over-ambitious. In the abstract itself, where it states, “the result could be largely affected by how ocean CO2 trends are obtained”, this doesn’t make sense because other DML techniques are not assigning or removing “ocean CO2 trends”. They are using atmospheric xCO2 as one (of many) driver variables and the machine learning approach estimates the pCO2 in areas missing data and from that ocean trends can be calculated. Therefore, this type of statement is not appropriate and especially not in the abstract.
Additionally, the statement in the abstract, “…and the sink in early and late year of the modelled period could be overestimates if CO2 trends were not well processed by models” is incorrect. Yes, this product has a smaller sink as compared to the others, but it does not show a difference in both early and late periods as compared to all other products- just those that use a fixed rate to correct observations to a base year. And what is “not well processed by models”? You must be careful about your terminology. Models are different from DML products as discussed in this and similar literature. So, to say that it is not well processed by models just doesn’t make any sense. The models are not processing the data- the machine learning methods are.
Regarding the comparison to other products, specifically Figure 6, first, the layout seems like a waste of space for these figures and it could be condensed more (also include a legend on each figure subplot for ease of interpretation). Second, the authors leave out a comparison to the JENA product which is one of the most long-available pCO2 products and has been included in the GCB for many years now. And lastly, they don’t account for the fact that the products all are calculating flux using their own choice of wind and atmospheric products, gas exchange parameterization and other choices. These choices could easily make the difference in the increase in flux seen in the last decade or so as compared to the presented NIES -ML3 product. For example, if the wind product chosen by CMEMS-FNN has stronger winds in the 2000s, it could result in a larger flux and therefore that would account for some of the difference seen here. Unless you are standardizing for all of these factors (as done in the pySeaFlux package/product ensemble) then you cannot confidently say the difference are due to their choice of method regarding co2 trend handling.
Another overarching question I had when reading through this method is that the authors need to be sure to adequately handle the fact that global location of the observations are totally variable in time and that definitely impacts the trend. There is no way to account for or correct for the fact that some years data is dominated from the tropical pacific which could have a much different trend than that of the rest of the ocean. This could play a large role in the differences seen as well.
The effort that the authors put to explain their choice of rate and how it impacts the model result is substantial and accounts for about half of the manuscript. But in the end, they get a value that is very close to the (smoothed) atmospheric xco2 increasing rate. The comparisons made in Figure 2 are great at showing a single trend is inappropriate for such a long-time scale of data that is now available, but it doesn’t show why this method of rate assessment is an improved method over just using atmospheric xCO2 trends.
In Section 3.2 where the flux is presented, the authors report an uncertainty value. It is important to note that this uncertainty is solely due to the machine learning choice used, and represents a spread around the three methods shown here. There is no uncertainty included from other aspects of the flux calculation such as the flux parameterization for example.
The flux convention used in this manuscript is opposite of that typically used in the observation-based pCO2 product literature. First and formost though, the convention is never defined and explained that a negative flux value would mean an efflux of carbon out of the ocean. I am confused as to why the authors chose this convention in the first place, especially then used a colorbar that is opposite of expected (negative values in warm colors) so that the maps look like the maps in the literature presenting other products. I would highly suggest flipping the convention around, specifically for the maps and discussion in section 3.2.
In regard to Figure 6 and the comparison to other available DML products, it could be that the differences seen in the later years could be influenced by the nearly logarithmic increase in available data for recent decades as opposed to the 1980s and 1990s. I would be interested to hear the authors comment son this in the discussion in Section 3.3.
Specific comments
There are way too many grammatical and clarifying errors to correct and list here. Some common ones are the use of the word “residue” where it likely should be “residual” and “special” where they meant “spatial”.
Line 149: “The YON was about 20 years in early 1990s” as shown in Figure 1b- I don’t see that from Figure 1b at all. It seems that the line has flattened well before 20 years on the x-axis. What makes the author state this value here?
Line 152: An example of where the authors are over-confident in their statements. “One of the reasons must be that the data points after 2000….” Is that really the ONLY reason that could possibly explain this? I don’t think so. A better phrase would be that one of the reasons “could be”…
Line 194: Figure 3 shows the annual fluxes and Figure 4 shows the spatial distributions.
Line 195: Typo- should be Global Carbon Budget 2021.
Figure 4: Include an explanation of what a, b, and c are in the figure caption. If you are going to state that “the patterns agree well with those in GCB-2021, then show this comparison. What models is this comparing to, and which compare well? The models, spatially, actually have a LARGE range of estimates of flux (and thus pCO2), see Fay & McKinley 2021 in GRL. To broadly say that these maps compare well with the GCB2021 models is untrue and unsupported in this paper.
Line 207: Why are you comparing your 1980-2020 to GCB-2021 for years 2011-2020? You show in Figure 6 that these are years where many of the products start to diverge. Why not do the comparison for the same subset of years?
Line 560: “GSB-2021” should be “GCB-2021”.
Citation: https://doi.org/10.5194/essd-2022-71-RC3 -
AC3: 'Reply on RC3', J. Zeng, 18 Apr 2022
The reader has given valuable comments and kindly pointed out many misuses of terminology, grammar and spelling errors. We have corrected the errors and incorporated most advice in the revised manuscript. The followings are point-to-point response to the comments.
- The reader suggested explaining more about the choice outlined 2.1 and 2.3. We have provided python code for model creation in the supplement. Those interested in the details could go on reading related python documents and references listed in the manuscript. In the revised manual script, we added more explanations for the iteration method of rate extraction.
- Regarding the question of why not just using atmospheric xCO2 trends. It sounds simple but is questionable. First, what xCO2 trends to use? Using the annual increase rate of xCO2 in the same year or in the previous year, or the mean rate of the past two years, three years… 10 years? Second, how to prove that ocean CO2 change follows the same pattern as xCO2? The uptake of air CO2 may not the only major factor that leads to ocean CO2 increase. Changing SST could be an important factor as well, for example. For these reasons, we do not plan to use xCO2 trends directly to model flux for comparison.
- We have rewritten the discussions for comparison. First, we included the JENA product. We didn’t include it initially because its spatial and temporal resolutions are different from others. The results of a comparison could depend on how a monthly product in 1x1 degree grids were derived from JENA’s daily product in 2x2.5 degree grids. Seconds, we recalculated the fluxes of all products under comparison using the same method and adjusted the fluxes to the ones as if they have the same grid coverage. Third, time-series fluxes are put in the same plot for comparison. Fourth, we added a figure to summaries the mean differences.
- The reader questioned the variable of global locations of observations. As the reader pointed out “There is no way to account for or correct for the fact that some years data is dominated from the tropical pacific which could have a much different trend than that of the rest of the ocean.” Our method examined the change of APPARENT GLOBAL TREND with data length and select the trend when the change become small. The implicit assumption is that the effect of unbalanced sampling becomes acceptably week with increasing data length. The method is not perfect, but it offers an alternative. If the locations were quite even geographically, it would not be necessary to use machine learning in the first place, as a simple spatial interpolation would be sufficient.
- In the revised manuscript, we emphasize that the uncertainty value in Section 3.2 is solely due to the machine learning. It is unreality to count in all uncertainty factors in such a data paper, including uncertainties of measurements, griding, and etc.
- In the revised manuscript, we explained that a negative flux value means an efflux of carbon out the ocean and a positive flux into the ocean. Although the reader strongly suggests flipping the convention around, we think the convection makes it easier to see the trend of flux. When we made the flux map, we used the same convention and a similar colorbar palette as those in the figure 6 of GCB-2021.
Citation: https://doi.org/10.5194/essd-2022-71-AC3
-
AC3: 'Reply on RC3', J. Zeng, 18 Apr 2022
-
RC4: 'Comment on essd-2022-71', Anonymous Referee #3, 28 Apr 2022
Review for ‘A new estimate of oceanic CO2 fluxes by machine learning reveals the
impact of CO2 trends in different methods’ by Jiye Zeng et alThe paper aims to investigate issues related to interpolating sparse in situ ocean pCO2 data to allow complete fields to be generated and then global assessments of air-sea gas fluxes and the net oceanic sink to be assessed. The study investigates the impact of assumed or specified trends that exist within some of these interpolation methods towards understanding what impact these assumed trends can have on the resultant oceanic net sink estimate (and in particular the results at the ends of time series). The analysis includes producing a novel reconstruction method which produces an integrated net sink value which is markedly different in magnitude to all of the other datasets analysed.
I would recommend that the paper is significantly revised, including a modified title before then being reviewed again. The reasons for this recommendation are given below.
My main comments are:
- The paper seems to be focussed on a trend analysis, and identifying the impact of assumed trends within different datasets, rather than providing ‘a new estimate of oceanic CO2 fluxes’ and determining the absolute oceanic sink. The methods identify how the rate of ocean absorption has changes and how each of the methods that are also compared (within the inter-comparison) does or does not capture similar trends. As a result the paper title seems a bit misleading as the title should really explain that's its focussed on the analysis of the trends and the rate of ocean absorption of carbon, and not the oceanic sink strength and magnitude (and not on presenting a new estimate of the oceanic sink). The focus in the title and abstract also seems to be a bit out of scope for the journal ESSD as I thought that the journal is focussed on providing datasets for wide use, where the focus here seems to be more focussed on novel science or result-focussed work.
- A detailed methods section seems to be missing from the paper. This lack of information makes it hard to evaluate the approach presented and the results. The limited amount of information on methods (which appear in the section called ‘data’) lack detail and need significantly expanding. In particular the methods used to calculate the air-sea fluxes and the derivation of each of the parameters needed as input to the flux calculation are missing (eg do the NOAAs Marine Boundary Layer Reference provide pCO2atmosphere data, or do they provide XCO2 data that you use to calculate pCO2atmosphere?) The calculation of the net integrated air-sea fluxes is also missing (eg how were the values in Figure 6 calculated, what land and ocean masks were used, how were mixed pixels containing land and ocean handled, was surface area calculated using assuming a sphere or an ellipsoid?). Providing a separate section on methods that detail all of the methods used would help the reader. I could not follow the methods in section 2.3 Rate extraction. A more detailed overview is really needed in the paper for this part of the methods (The authors rely on the reviewer/reader reading another paper to gain a basic understanding).
- The difference between air-sea flux data products is likely to be at least partly (maybe mostly) driven by the differences in input data for calculating the air-sea gas fluxes (eg differences in wind speed data, or sea surface temperature, salinity etc) and this could also be one of the causes for the different identified trends. This issue has been overlooked and should at least be discussed and/or tested for. It may be possible to partially remove it, eg by using a common method to calculate the gas fluxes, a common set of wind data, common atmospheric data and then just the pCO2water fields from each of the datasets within the inter-comparison. This would allow the authors to isolate the impact of different wind and atmospheric forcing data.
- All of the spatially complete datasets used within the inter-comparison form submissions to the Global Carbon Budget assessments. Within the GCB assessments these datasets are all assessed using SOCAT data. If the authors want to keep the ‘new estimate of oceanic CO2 fluxes’ as a key finding and focus of the paper, then their new pCO2 fields should be evaluated using the community SOCAT database eg train on a subset of the dataset, test against the remainder (eg RMSD and bias), or at the very least evaluate their complete pCO2 dataset using the data that were used to train the approach. At the moment the paper presents an inter-comparison where the new approach is vastly different from all other data included in the inter-comparison but no justification or evaluation as to why these new data are useful is given; only that that they are different and new. Can the authors provide some sort of statistical assessment or evaluation of the output data (pCO2 fields and/or gas fluxes) to provide the reader with some confidence that the new results are valid?
General comments
The paper would benefit from the help of a copy editor help improve the main text.
Citation: https://doi.org/10.5194/essd-2022-71-RC4 -
AC4: 'Reply on RC4', J. Zeng, 02 May 2022
The reviewer has given valuable opinions, including changing the title to focus on the dataset. Yes, our intention is to improve our previous method for ocean CO2 reconstruction, not the focus on scientific discovery. We are considering to change the title to “A surface oceanic CO2 product reconstructed by using machine learning to extract CO2 trends at decadal scale”. The following are point-to-point to the reviewer’s comments.
- We are adding more details in the method section to address the issues of air-sea flux calculation, rate extraction, and etc.
- Instead of using flux values in comparison products, we used their pCO2 to recalculate fluxes by a common method. Discussions on the comparison have been revised substantially.
- Our method did set aside part of SOCAT fCO2 for validation. We called it “leave-one-year-out” validation method. For each of the three machine learning methods, 41 validations were done by set aside data in 1980 to 2020 year by year for validation and use the rest for training.
Citation: https://doi.org/10.5194/essd-2022-71-AC4
Data sets
NIES-ML3 ensemble product of surface ocean CO2 concentrations and air-sea CO2 fluxes reconstructed by using three machine learning models with new CO2 trends Jiye Zeng https://db.cger.nies.go.jp/DL/10.17595/20220311.001.html.en
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 822 | 360 | 45 | 1,227 | 73 | 33 | 37 |
- HTML: 822
- PDF: 360
- XML: 45
- Total: 1,227
- Supplement: 73
- BibTeX: 33
- EndNote: 37
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1